From 59c47dcb719f0ed7c4d7364a332732065aaa0712 Mon Sep 17 00:00:00 2001
From: LauraGPT <18321252+LauraGPT@users.noreply.github.com>
Date: Mon, 31 Aug 2026 07:55:07 +0000
Subject: [PATCH] docs: keep README updates focused

Signed-off-by: LauraGPT <18321252+LauraGPT@users.noreply.github.com>
---
 README.md                                  | 57 ++--------------------
 README_ja.md                               | 36 +++-----------
 README_ko.md                               | 36 +++-----------
 README_zh.md                               | 54 ++------------------
 tests/test_docs_funasr_install_commands.py | 19 +++++++-
 5 files changed, 39 insertions(+), 163 deletions(-)

diff --git a/README.md b/README.md
index 28c3bd4ed..4a2d6bbbc 100644
--- a/README.md
+++ b/README.md
@@ -333,59 +333,12 @@ inference on physical Blackwell hardware.
 
 ## What's new
 
-- 2026/08/31: **v1.4.11 on PyPI** — preserves punctuation in mixed-language readable subtitles when a timestamped ASR word carries the SentencePiece `▁` boundary marker. Previously one invisible marker could invalidate punctuation-to-timestamp alignment for the whole recording and fall back to punctuation-free VAD segments even though the top-level transcript was punctuated. On the reporter's 10m33s sample, all 272 predicted punctuation marks now reach 275 timestamped sentences and 137 readable cues, with no cue over 7.94 seconds or 42 characters; the normal comparison sample remains aligned. Upgrade with `python -m pip install -U "funasr==1.4.11"`; [#3539](https://github.com/modelscope/FunASR/issues/3539) remains open pending reporter confirmation. [Fix ->](https://github.com/modelscope/FunASR/pull/3587) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.11)
-- 2026/08/31: **v1.4.10 on PyPI** — preserves readable Chinese word boundaries when an overlong subtitle sentence must be split. Instead of filling every cue to the hard character limit, the fallback balances duration and length while preferring Jieba lexical boundaries, punctuation, script transitions, and real timestamp gaps. On the reporter's exact 10m33s mixed Chinese/Korean sample, 84 source segments became 133 cues with no cue over 8 seconds or 42 characters, while phrases such as `钟书成为` and `扭了一下` remain intact. Upgrade with `python -m pip install -U "funasr==1.4.10"`; [#3539](https://github.com/modelscope/FunASR/issues/3539) remains open pending reporter confirmation. [Fix ->](https://github.com/modelscope/FunASR/pull/3583) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.10)
-- 2026/08/30: **v1.4.9 on PyPI** — makes the subtitle fix merged after v1.4.8 directly installable. Readable SRT rendering now carries aligned word timestamps into each source sentence and splits an already overlong sentence only at real model time boundaries. On the reporter's 10m33s mixed Chinese/Korean sample, the exact merged code changed 84 cues with 26 over eight seconds into 131 cues with none over eight seconds or 42 characters, without overlaps or fabricated timestamps. Upgrade with `python -m pip install -U "funasr==1.4.9"`; [#3539](https://github.com/modelscope/FunASR/issues/3539) remains open until the reporter retests. [Fix ->](https://github.com/modelscope/FunASR/pull/3574) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.9)
-- 2026/08/30: **v1.4.8 on PyPI** — completes the third-party MOSS-Transcribe-Diarize vLLM path added after v1.4.7. `backend="vllm"` now accepts the official `response_format=diarized_json` speaker segments and normalizes them into FunASR `sentence_info`; `max_completion_tokens` is forwarded for long meetings, with 8192 tokens documented for recordings that exceed the default 5120-token boundary. MOSS still performs transcription, timestamps, and diarization end to end, so do not add external `vad_model` or `spk_model`. Upgrade with `python -m pip install -U "funasr==1.4.8"`. The GitHub Release pairs this package with the verified ten-platform llama.cpp v0.2.6 archives. [MOSS deployment guide ->](./docs/moss_transcribe_diarize.md) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.8)
-- 2026/08/30: **llama.cpp runtime v0.2.6** — adds a dedicated Windows CUDA architecture 120 (`sm_120`) archive for RTX 50 / Blackwell GPUs while retaining the architecture 86 package. The two CUDA ZIPs bundle the required NVIDIA cuBLAS DLLs and license, statically link the MSVC runtime, and pass PE import audits; all ten Linux, macOS, and Windows archives are built from one exact release commit and checked against their published SHA-256 values. Build and package verification does not prove successful inference on physical Blackwell hardware, so hardware reports remain open until users retest the matching archive. [Implementation ->](https://github.com/modelscope/FunASR/pull/3570) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6)
-- 2026/08/30: **v1.4.7 on PyPI** — integrates the third-party OpenMOSS MOSS-Transcribe-Diarize model with FunASR `AutoModel`. Choose local Transformers (`backend="hf"`) or an existing vLLM server (`backend="vllm"`); both paths normalize the tagged model output into `text`, `raw_text`, millisecond `timestamp`, and speaker-aware `sentence_info`. MOSS performs transcription, timestamps, and diarization in one pass, so omit external `vad_model` and `spk_model`. This release also improves SRT cue continuity across tiny same-speaker gaps and adds opt-in realtime decode profiling for diagnosing long-session latency. Upgrade with `python -m pip install -U "funasr==1.4.7"`. The GitHub Release carries the verified nine-platform llama.cpp v0.2.5 runtime archives. [MOSS deployment guide ->](./docs/moss_transcribe_diarize.md) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.7)
-- 2026/08/29: **llama.cpp runtime v0.2.5** — uploads host weights into the selected Vulkan backend buffer before graph execution. Q8 and F16 weights pass local Linux Vulkan llvmpipe validation, and all nine Linux, macOS, and Windows archives were built from exact commit `f371370d4c5e4c61d13d4eb9c55cda2f4dd95e4f` and verified against their published SHA-256 values. This does not claim the AMD Windows hardware crash is fixed; [#3479](https://github.com/modelscope/FunASR/issues/3479) remains open for reporter retesting. [Fix ->](https://github.com/modelscope/FunASR/pull/3555) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.5)
-- 2026/08/29: **llama.cpp runtime v0.2.4** — fixes intermittent empty SenseVoiceSmall transcripts with F16 GGUF models. Query embeddings are now decoded according to their GGML F16/F32 type instead of reading F16 storage as F32. The exact v0.2.3 AVX2 asset reproduced 22 blank outputs across 298 completed runs; the fixed head produced one identical non-empty result in 100/100 runs and matched the Q8 model byte-for-byte. The release workflow will build nine Linux, macOS, and Windows archives from one exact commit. [Fix ->](https://github.com/modelscope/FunASR/pull/3550) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.4)
-- 2026/08/29: **v1.4.6 on PyPI** — the realtime WebSocket server keeps 20-second pings but no longer closes healthy sessions by default when queued decoding delays pong handling; operators can still configure a positive timeout after measuring production queue and decode latency. This release also avoids `IndexError` on empty English timestamp segments and groups adjacent subtitle words into readable cues. Upgrade with `python -m pip install -U "funasr==1.4.6"`. The GitHub Release carries the verified nine-platform llama.cpp v0.2.3 runtime archives. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.6)
-- 2026/08/29: **llama.cpp runtime v0.2.3** — adds flushed stage boundaries after backend initialization for model loading, audio/VAD, graph construction and allocation, and compute. These diagnostics narrow Windows AMD Vulkan `0xC0000005` failures that occur after `vulkan backend ready`; they do not claim the hardware-specific crash is fixed. Nine Linux, macOS, and Windows archives are published from one exact commit. [Troubleshooting boundaries ->](./runtime/llama.cpp/README.md#optional-windows-vulkan-backend-for-sensevoicesmall) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.3)
-- 2026/08/28: **v1.4.5 on PyPI** — `torchaudio` is no longer a hard inference dependency. Feature extraction uses `torchaudio.compliance.kaldi` when available or the optional `kaldi-native-fbank` backend, while operations that still require `torchaudio` now fail with an actionable message. The fallback was verified end to end on Ascend 910B: 70.47 s of audio processed in 1.15 s (RTF 0.016). Upgrade with `python -m pip install -U "funasr==1.4.5"`; on systems without `torchaudio`, install `python -m pip install -U "funasr[knf]==1.4.5"`. The GitHub Release also carries the verified nine-platform llama.cpp v0.2.1 runtime archives. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.5)
-- 2026/08/27: **llama.cpp runtime v0.2.1** — Vulkan device selection now accepts matching integrated GPUs, prefers a matching discrete GPU when both are available, and otherwise falls back to the iGPU. Nine Linux, macOS, and Windows archives were rebuilt and verified with published SHA-256 values. Radeon 780M still needs reporter hardware confirmation; the separate RX 9070 XT `0xC0000005` initialization crash is not claimed fixed. [Download matrix & quickstart ->](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.1)
-- 2026/08/26: **v1.4.4 on PyPI** — realtime WebSocket decoding now batches compatible sessions instead of serializing every connection behind one process-wide lock. On the H100 regression workload, 12-client STOP p95 fell from 19.8 s to 0.4 s and 16-client aggregate throughput rose from 8.6x to 13.2x, with no client errors. This hotfix also restores compatibility with PyTorch releases that do not expose `torch.amp` and raises real runtime-binding exceptions. Upgrade with `python -m pip install -U "funasr==1.4.4"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.4)
-- 2026/08/21: **v1.4.3 on PyPI** — `AutoModel(vad_model="silero-vad")` now offers an optional Silero VAD adapter with millisecond segments, configurable thresholds, 8/16 kHz input, ONNX mode, and bounded segment lengths. Upgrade with `python -m pip install -U "funasr==1.4.3"`; enable the adapter with `python -m pip install -U "funasr[silero]==1.4.3"`. Speaker diarization with a known speaker count now routes large embedding sets through fixed-K clustering instead of memory-heavy dense spectral clustering. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.3)
-- 2026/08/14: **v1.4.2 on PyPI** — punctuation-aware sentence alignment now preserves subtitle segmentation when a punctuation token boundary splits a timestamped ASR word. Distributed training now synchronizes DDP/FSDP gradients on the final microbatch of each accumulation window and initializes DeepSpeed/FSDP mode from the resolved configuration. The tagged source also includes llama.cpp SRT output and the v0.2.0 AMD Vulkan submission updates. Install with `python -m pip install -U "funasr==1.4.2"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.2)
-- 2026/08/11: **llama.cpp runtime v0.2.0** — pins upstream llama.cpp at `803b7fca` and publishes nine SHA-256-listed Linux, macOS, and Windows archives from one tested workflow. Fun-ASR-Nano, SenseVoice, and Paraformer CLIs can now write SRT subtitles; Vulkan startup reports actionable AMD diagnostics and a CPU fallback. The AMD Windows Vulkan crash fix still awaits confirmation on the reporter's hardware. [Download matrix & quickstart ->](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.0)
-- 2026/08/04: **v1.4.1 on PyPI** — the Hugging Face `paraformer-en` alias now resolves to the official English checkpoint instead of silently downloading the Chinese model. This patch also includes Fun-ASR-Nano LoRA fine-tuning and safer checkpoint handling; the tagged source adds JSONL timestamp output, SenseVoice TensorRT deployment, and the OpenClaw realtime transcription integration. Install with `python -m pip install -U "funasr==1.4.1"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.1)
-- 2026/08/04: **OpenClaw realtime transcription integration** — the new [`openclaw-funasr`](integrations/openclaw/) source package connects self-hosted FunASR `online`, `offline`, and `2pass` WebSocket recognition to OpenClaw Talk and Voice Call. Its 8 kHz G.711 mu-law conversion, 60 ms framing, partial/final transcripts, reconnect limits, package installation, and runtime registration have been verified against OpenClaw `2026.7.2`; npm and ClawHub publication will follow the required [upstream SDK change](https://github.com/openclaw/openclaw/pull/118977).
-- 2026/07/31: **v1.4.0 on PyPI** — `AutoModel` now rejects the common `vda_model` misspelling before model download and points to `vad_model`, so VAD-dependent segmentation, speaker processing, and `sentence_info` are not silently disabled. The GitHub source release also updates the legacy WebSocket file runtime: clients wait for an explicit end-of-input acknowledgement while the server flushes pending offline, online, and 2pass audio and reports finalization failures. Install the Python package with `python -m pip install -U "funasr==1.4.0"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.0)
-- 2026/07/27: **v1.3.30 on PyPI** — container-formatted WAV, MP3, FLAC, OGG, MP4/M4A, and WebM audio bytes are now decoded through their codecs instead of being misread as raw PCM. OpenAI-compatible responses preserve speaker labels, VAD sentence timing survives punctuation mismatch, trusted browser clients can opt in to CORS, and vLLM VAD chunks are capped at 30 seconds. The GitHub release also includes the current prebuilt llama.cpp runtime for nine desktop and server targets. Install with `python -m pip install -U "funasr==1.3.30"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.30)
-- 2026/07/24: **v1.3.29 hotfix on PyPI** — SenseVoice long-audio inference now returns each VAD speech region through `sentence_info` when token timestamps and a punctuation model are unavailable. Subtitle clients receive the recognized text with real millisecond start/end bounds instead of one zero-length or full-media cue. Install with `python -m pip install -U "funasr==1.3.29"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.29)
-- 2026/07/24: **v1.3.28 hotfix on PyPI** — realtime WebSocket finalization now preserves clean continuous partial transcripts when a VAD-locked decode truncates to a short prefix, repeats a hallucinated phrase, or raises; short STOP tails, VAD finalization, and speaker completion now share the same reliable path. SenseVoice subtitle segmentation also aligns rich tags, punctuation, and word/BPE timestamps without collapsing Chinese into one cue or damaging English surface text. Install with `python -m pip install -U "funasr==1.3.28"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.28)
-- 2026/07/24: **v1.3.27 on PyPI** — the OpenAI-compatible server now reports detected SenseVoice language metadata in `verbose_json` and reuses the cached Fun-ASR-Nano `AutoModel` after vLLM fallback. When vLLM/VAD setup and its fallback both fail, half-initialized engine state is cleared so a later request can retry. Install with `python -m pip install -U "funasr==1.3.27"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.27)
-- 2026/07/23: **llama.cpp runtime v0.1.9** — adds `funasr-llamacpp-windows-x64-vulkan.zip` for standalone SenseVoiceSmall Vulkan inference on Windows with AMD, Intel, or NVIDIA drivers. Linux Vulkan, Windows CUDA, CPU/AVX2, Linux arm64, and macOS arm64 assets remain available. [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.9)
-- 2026/07/23: **v1.3.26 on PyPI** — `funasr-server --model fun-asr-nano --hub ms` now honors the requested ModelScope hub for the default Fun-ASR-Nano model in both the vLLM path and the AutoModel fallback, avoiding unintended Hugging Face downloads when users choose ModelScope. Install with `python -m pip install -U "funasr==1.3.26"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.26)
-- 2026/07/23: **v1.3.25 on PyPI** — realtime WebSocket users can now use deterministic final-text hotword corrections with `POSTPROCESS_HOTWORDS:wrong=>right` or `--postprocess-hotword-file`, keeping fixed-name cleanup separate from model-level `HOTWORDS:` decoding bias. The source-tree realtime entrypoint also works without preinstalling the package. Install with `python -m pip install -U "funasr==1.3.25"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.25)
-- 2026/07/23: **v1.3.24 on PyPI** — OpenAI-compatible server deployments now support custom model paths and hub selection, the llama.cpp/GGUF runtime docs include the HTTP transcription wrapper and Linux Vulkan package, and public docs links were refreshed for cleaner onboarding. Install with `python -m pip install -U "funasr==1.3.24"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.24)
-- 2026/07/22: **v1.3.23 on PyPI** — packaging and onboarding refresh for this week's community integrations: the PyPI long description now highlights the current OpenAI-compatible server path, llama.cpp/GGUF runtime notes, Windows CUDA architecture guidance, and browser quickstart links shipped in the repository docs. Runtime code is unchanged from v1.3.22. Install with `python -m pip install -U "funasr==1.3.23"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.23)
-- 2026/07/22: **llama.cpp runtime v0.1.8** — adds `funasr-llamacpp-linux-x64-vulkan.tar.gz` for SenseVoiceSmall on Linux Vulkan GPUs. Run `llama-funasr-sensevoice ... --backend vulkan`; CPU, AVX2, macOS arm64, Windows CPU/AVX2, and Windows CUDA packages remain available. [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.8)
-- 2026/07/19: **v1.3.22 on PyPI** — `funasr-server` now fills OpenAI-compatible `verbose_json.segments` for text-only SenseVoice/Paraformer fallback responses, so subtitle clients no longer see an empty `segments` array when `text` is populated. Install with `python -m pip install -U "funasr==1.3.22"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.22)
-- 2026/07/19: **v1.3.21 on PyPI** — fixes first-import onboarding in fresh environments where users install `funasr` before choosing a platform-specific PyTorch build. `import funasr` and `funasr.__version__` now work without torch; accessing `AutoModel` still requires PyTorch and raises a clear install hint. Install with `python -m pip install -U "funasr==1.3.21"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.21)
-- 2026/07/19: **v1.3.20 on PyPI** — PyPI metadata and install guidance now point at the current FunASR docs, community integrations, and quoted `python -m pip install -U "funasr>=1.3.19"` commands for Fun-ASR-Nano deployment paths. This is a documentation/packaging sync; runtime code remains unchanged from v1.3.19. Install with `python -m pip install -U "funasr==1.3.20"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.20)
-- 2026/07/19: **v1.3.19 on PyPI** — realtime WebSocket long-session troubleshooting docs are now shipped with the package. Run the server with `--enable-spk --log-session-stats-interval 30` and attach the emitted `Session stats:` lines when reporting disconnects or memory growth. Install with `python -m pip install -U "funasr==1.3.19"`. [Long-session guide ->](docs/vllm_guide.md#long-session-diagnostics) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.19)
-- 2026/07/19: **v1.3.18 on PyPI** — CLI SRT/TSV subtitle output now requests sentence timestamps and loads punctuation when needed, so `funasr audio.wav --output-format srt --output-dir ./subs` writes segmented subtitle cues instead of one full-text block. Install with `python -m pip install -U "funasr==1.3.18"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.3.18)
-- 2026/07/18: **v1.3.16 on PyPI** — client-driven realtime endpoints for Fun-ASR-Nano. Start one WebSocket session, stream PCM, and send `COMMIT` for each utterance without loading server-side VAD; short utterances finalize and timestamps remain monotonic across commits. Install with `pip install --upgrade funasr`, then run `funasr-realtime-server --endpoint-mode client`. [Guide →](examples/industrial_data_pretraining/fun_asr_nano/docs/realtime_demo.md)
-- 2026/07/18: **llama.cpp runtime v0.1.7** — prebuilt Windows CUDA package for SenseVoiceSmall (`funasr-llamacpp-windows-x64-cuda.zip`) plus Linux / macOS / Windows CPU packages. Download the GGUF model, then run `llama-funasr-sensevoice ... --backend cuda` on supported NVIDIA GPUs. [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.7)
-- 2026/06/20: **llama.cpp / GGUF runtime** — run SenseVoice / Paraformer / Fun-ASR-Nano on CPU & edge as a single self-contained binary (a whisper.cpp-style alternative), built-in FSMN-VAD, no Python at runtime. Prebuilt binaries for Linux / macOS / Windows + **q8 quantized models (~half the size, same accuracy)**. [runtime/llama.cpp/](./runtime/llama.cpp/) · [Releases](https://github.com/modelscope/FunASR/releases)
-- 2026/06/21: **v1.3.12** on PyPI — rolling fixes (qwen3-asr language codes, glm_asr, vLLM repetition_penalty). `pip install --upgrade funasr`
-- 2026/05/24: **vLLM Inference Engine** — 2-3x faster LLM decoding for Fun-ASR-Nano. Streaming WebSocket service with VAD + Speaker Diarization. [Guide →](docs/vllm_guide.md) · [Realtime WS tuning →](docs/vllm_guide.md#67-production-concurrency-and-multi-process-deployment) · [API stability checklist →](docs/vllm_guide.md#production-api-stability-checklist)
-- 2026/05/24: **Dynamic VAD** — adaptive silence threshold (default on). Short sentences stay intact, long segments get auto-split. [Details →](docs/vllm_guide.md#附录dynamicstreamingvad)
-- 2026/05/24: **v1.3.3** — `funasr-server` CLI, OpenAI-compatible API, MCP Server for AI agents. `pip install --upgrade funasr`
-- 2026/05/20: Added Qwen3-ASR (0.6B/1.7B) — 52 languages, auto detection. [usage](examples/industrial_data_pretraining/qwen3_asr)
-- 2026/05/20: Added GLM-ASR-Nano (1.5B) — 17 languages, dialect support. [usage](examples/industrial_data_pretraining/glm_asr)
-- 2026/05/19: Fun-ASR-Nano and SenseVoice can be combined with VAD and CAM++ for speaker diarization.
-- 2025/12/15: [Fun-ASR-Nano-2512](https://github.com/QwenAudio/Fun-ASR) — Chinese, English, Japanese, and Chinese dialect support; trained on tens of millions of hours.
-
-<details><summary>Older</summary>
-
-- 2024/10/10: Whisper-large-v3-turbo support added.
-- 2024/07/04: [SenseVoice](https://github.com/QwenAudio/SenseVoice) — ASR + emotion + audio events.
-- 2024/01/30: FunASR 1.0 released.
+- **FunASR 1.4.11** is the current PyPI release. It improves readable multilingual subtitles while preserving real model timestamps. Install with `python -m pip install -U "funasr==1.4.11"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.11)
+- **MOSS-Transcribe-Diarize integration** adds long-form ASR, timestamps, and anonymous speaker labels in one generation, without an external VAD or speaker pipeline. FunASR supports local Transformers and existing vLLM/SGLang services; FunClip can export speaker-aware SRT and clips. [Deployment guide ->](./docs/moss_transcribe_diarize.md)
+- **llama.cpp runtime v0.2.6** provides verified prebuilt archives for ten Linux, macOS, and Windows targets, including dedicated Windows CUDA packages for RTX 30/40 and RTX 50 GPUs. [Download matrix ->](https://www.funasr.com/en/deploy/llama-cpp.html)
+- **Realtime serving is faster and more resilient**: compatible WebSocket sessions are batched instead of serialized, and queued decoding no longer closes healthy sessions by default. On the H100 regression workload, 12-client STOP p95 fell from 19.8 s to 0.4 s. [Production guide ->](./docs/vllm_guide.md)
 
-</details>
+> See [GitHub Releases](https://github.com/modelscope/FunASR/releases) for the complete changelog and downloadable assets.
 
 ---
 
diff --git a/README_ja.md b/README_ja.md
index 99039238a..36a2ae103 100644
--- a/README_ja.md
+++ b/README_ja.md
@@ -99,36 +99,12 @@ Whisper は単一モデルですが、**FunASR はツールキット**です—
 
 ## 最新情報
 
-- 2026/08/31：**v1.4.11 を PyPI に公開** — timestamp 付き ASR word に SentencePiece の `▁` boundary marker が含まれると、多言語 readable subtitle の punctuation が失われる問題を修正しました。以前は一つの不可視 marker だけで録音全体の punctuation-to-timestamp alignment が失敗し、top-level transcript に punctuation があっても punctuation-free VAD segment に fallback していました。Reporter の 10分33秒 sample では、予測された 272 punctuation が 275 timestamped sentences と 137 readable cues にすべて保持され、最大 7.94秒・42文字です。正常な comparison sample の alignment も維持しています。更新：`python -m pip install -U "funasr==1.4.11"`。[#3539](https://github.com/modelscope/FunASR/issues/3539) は reporter の確認まで open のままです。[Fix ->](https://github.com/modelscope/FunASR/pull/3587) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.11)
-- 2026/08/31：**v1.4.10 を PyPI に公開** — 長すぎる字幕を分割するときに中国語の単語境界を壊す問題を修正しました。Fallback は文字数上限まで機械的に詰めず、8秒・42文字の hard limit 内で duration と length を均衡させ、Jieba lexical boundary、punctuation、script transition、実 timestamp gap を優先します。Reporter と同じ 10分33秒の中国語・韓国語 mixed sample では、84 source segments が 133 cues になり、全 cue が制限内で、`钟书成为` と `扭了一下` などの phrase も保持されました。更新：`python -m pip install -U "funasr==1.4.10"`。[#3539](https://github.com/modelscope/FunASR/issues/3539) は reporter の確認まで open のままです。[Fix ->](https://github.com/modelscope/FunASR/pull/3583) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.10)
-- 2026/08/30：**v1.4.9 を PyPI に公開** — v1.4.8 公開後に merge された subtitle 修正を直接 install できる release にしました。Readable SRT は aligned word timestamp を source sentence に保持し、すでに長すぎる sentence を model の実 timestamp boundary だけで分割します。Reporter の 10分33秒の中国語・韓国語 mixed sample では、84 cues（8秒超 26 cues）が 131 cues になり、8秒または 42文字を超える cue、overlap、fabricated timestamp は 0 でした。更新：`python -m pip install -U "funasr==1.4.9"`。[#3539](https://github.com/modelscope/FunASR/issues/3539) は reporter の再テストまで open のままです。[Fix ->](https://github.com/modelscope/FunASR/pull/3574) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.9)
-- 2026/08/30：**v1.4.8 を PyPI に公開** — v1.4.7 以降に merge された third-party MOSS-Transcribe-Diarize の vLLM path を完成させました。`backend="vllm"` は公式の `response_format=diarized_json` speaker segment を受け取り、FunASR `sentence_info` に正規化します。長時間会議では `max_completion_tokens` を server に渡せ、default の 5120-token 境界を超える録音向けに 8192 tokens を案内します。MOSS は transcription、timestamp、speaker diarization を end-to-end で実行するため、外部の `vad_model` や `spk_model` は追加しません。更新：`python -m pip install -U "funasr==1.4.8"`。GitHub Release は検証済み llama.cpp v0.2.6 の 10 platform archive と組み合わせます。[MOSS guide ->](./docs/moss_transcribe_diarize.md) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.8)
-- 2026/08/30：**llama.cpp runtime v0.2.6** — RTX 50 / Blackwell 向けに Windows CUDA architecture 120（`sm_120`）専用 archive を追加し、architecture 86 package も維持します。2 つの CUDA ZIP は必要な NVIDIA cuBLAS DLL と license を同梱し、MSVC runtime を static link して PE import audit に合格しています。Linux、macOS、Windows 向け 10 archive は同一の exact release commit から build し、公開 SHA-256 と照合します。Build と package の検証は Blackwell 実機 inference の成功を証明しないため、hardware report は対応 archive の再テストまで open のままです。[実装 →](https://github.com/modelscope/FunASR/pull/3570) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6)
-- 2026/08/30：**v1.4.7 を PyPI に公開** — OpenMOSS の third-party model MOSS-Transcribe-Diarize を FunASR `AutoModel` に統合しました。local Transformers（`backend="hf"`）または既存の vLLM server（`backend="vllm"`）を選択でき、どちらも tag 付き出力を `text`、`raw_text`、millisecond `timestamp`、speaker-aware `sentence_info` に正規化します。MOSS は transcription、timestamp、speaker diarization を 1 回の inference で実行するため、外部の `vad_model` と `spk_model` は指定しません。さらに、同一 speaker のごく短い gap をまたぐ SRT cue の連続性を改善し、long session latency の診断用に opt-in realtime decode profiling を追加しました。更新：`python -m pip install -U "funasr==1.4.7"`。GitHub Release には検証済みの llama.cpp v0.2.5 9 platform runtime archive も含まれます。[MOSS guide →](./docs/moss_transcribe_diarize.md) · [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.7)
-- 2026/08/29：**llama.cpp runtime v0.2.5** — graph 実行前に host weight を選択された Vulkan backend buffer へ upload します。Q8/F16 weight は Linux Vulkan llvmpipe で検証済みで、Linux、macOS、Windows 向け 9 archive は exact commit `f371370d4c5e4c61d13d4eb9c55cda2f4dd95e4f` から build し、公開 SHA-256 と照合しました。AMD Windows hardware crash の修正済みとはしておらず、[#3479](https://github.com/modelscope/FunASR/issues/3479) は reporter の実機再テスト待ちで open のままです。[修正 →](https://github.com/modelscope/FunASR/pull/3555) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.5)
-- 2026/08/29：**llama.cpp runtime v0.2.4** — F16 GGUF model で transcript が断続的に空になる問題を修正しました。Query embedding は F16 storage を F32 として読むのではなく、GGML の F16/F32 type に従って decode されます。正確な v0.2.3 AVX2 asset では 298 回の完了 run 中 22 回の空出力を再現しましたが、修正後は 100/100 回で同一の非空出力となり、Q8 model と byte 単位で一致しました。Release workflow は同一の exact commit から 9 platform archive を構築します。[修正 →](https://github.com/modelscope/FunASR/pull/3550) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.4)
-- 2026/08/29：**v1.4.6 を PyPI に公開** — realtime WebSocket server は 20 秒間隔の ping を維持しつつ、decode queue によって pong 処理が遅れた正常な session を default では切断しなくなりました。運用環境の queue と decode latency を測定したうえで、正の timeout を明示設定することもできます。空の英語 timestamp segment による `IndexError` を防ぎ、隣接する subtitle word を読みやすい cue にまとめます。更新：`python -m pip install -U "funasr==1.4.6"`。GitHub Release には検証済みの llama.cpp v0.2.3 9 platform runtime archive も含まれます。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.6)
-- 2026/08/29：**llama.cpp runtime v0.2.3** — backend 初期化後に、model 読み込み、audio/VAD、graph の構築と割り当て、compute の各段階を即時 flush する境界ログを追加しました。`vulkan backend ready` 後に起きる Windows AMD Vulkan `0xC0000005` の発生箇所を絞り込むための診断であり、hardware 固有の crash が修正済みとはしていません。同一の exact commit から Linux、macOS、Windows 向け 9 archive を公開します。[トラブルシューティング →](./runtime/llama.cpp/README.md#optional-windows-vulkan-backend-for-sensevoicesmall) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.3)
-- 2026/08/28：**v1.4.5 を PyPI に公開** — `torchaudio` は推論の必須依存関係ではなくなりました。特徴抽出は利用可能な場合に `torchaudio.compliance.kaldi` を使い、そうでなければ任意の `kaldi-native-fbank` backend を利用できます。引き続き `torchaudio` が必要な処理は、実行可能な依存関係メッセージを返します。この fallback は Ascend 910B で end-to-end 検証済みで、70.47 秒の音声を 1.15 秒（RTF 0.016）で処理しました。更新：`python -m pip install -U "funasr==1.4.5"`。`torchaudio` がない環境：`python -m pip install -U "funasr[knf]==1.4.5"`。GitHub Release には検証済みの llama.cpp v0.2.1 9 platform runtime archive も含まれます。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.5)
-- 2026/08/27：**llama.cpp runtime v0.2.1** — Vulkan のデバイス選択が一致する統合 GPU を受け入れ、一致する discrete GPU があればそちらを優先し、なければ iGPU にフォールバックするようになりました。SHA-256 を公開した Linux、macOS、Windows 向け 9 archive を再構築・検証済みです。Radeon 780M は報告者の実機確認待ちで、別件の RX 9070 XT `0xC0000005` 初期化 crash は修正済みとはしていません。[ダウンロード一覧と quickstart →](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.1)
-- 2026/08/26：**v1.4.4 を PyPI に公開** — realtime WebSocket decode は、全 connection を 1 つの process-wide lock で直列化せず、互換性のある session を batch 処理するようになりました。H100 regression workload では、12 client の STOP p95 が 19.8 秒から 0.4 秒に短縮され、16 client の aggregate throughput は 8.6x から 13.2x に向上し、client error は 0 件でした。この hotfix は `torch.amp` を持たない PyTorch release との互換性も復元し、runtime binding から実際の例外を送出します。更新：`python -m pip install -U "funasr==1.4.4"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.4)
-- 2026/08/21：**v1.4.3 を PyPI に公開** — `AutoModel(vad_model="silero-vad")` で任意の Silero VAD アダプターを利用でき、ミリ秒単位の区間、しきい値、8/16 kHz 入力、ONNX mode、最大区間長を設定できます。更新：`python -m pip install -U "funasr==1.4.3"`。アダプターの有効化：`python -m pip install -U "funasr[silero]==1.4.3"`。話者数が既知の話者ダイアライゼーションは、大規模 embedding をメモリ負荷の高い密なスペクトラルクラスタリングではなく fixed-K クラスタリングで処理します。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.3)
-- 2026/08/14：**v1.4.2 を PyPI に公開** — 句読点モデルの token 境界がタイムスタンプ付き ASR 単語の内部にある場合でも、文アラインメントが字幕分割を正しく保持するようになりました。分散学習では、各勾配累積 window の最後の microbatch で DDP/FSDP 勾配を同期し、解決済み設定から DeepSpeed/FSDP mode を初期化します。対応する GitHub ソース tag には llama.cpp SRT 出力と v0.2.0 の AMD Vulkan submission 更新も含まれます。インストール：`python -m pip install -U "funasr==1.4.2"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.2)
-- 2026/08/11：**llama.cpp runtime v0.2.0** — upstream llama.cpp を `803b7fca` に固定し、同一のテスト済み workflow から SHA-256 付きの Linux、macOS、Windows 向け 9 archive を公開しました。Fun-ASR-Nano、SenseVoice、Paraformer CLI は SRT 字幕を出力でき、Vulkan 起動時には AMD 向け診断と CPU fallback を案内します。AMD Windows Vulkan crash の修正は、報告者の実機確認待ちです。[ダウンロード一覧と quickstart →](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.0)
-- 2026/08/04：**v1.4.1 を PyPI に公開** — Hugging Face の `paraformer-en` エイリアスが、誤って中国語モデルをダウンロードせず、公式の英語 checkpoint を解決するようになりました。このパッチには Fun-ASR-Nano の LoRA 微調整と、より安全な checkpoint 処理も含まれます。対応する GitHub ソース tag には JSONL タイムスタンプ出力、SenseVoice TensorRT デプロイ、OpenClaw リアルタイム文字起こし連携も含まれます。インストール：`python -m pip install -U "funasr==1.4.1"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.1)
-- 2026/07/31：**v1.4.0 を PyPI に公開** — `AutoModel` はモデルのダウンロード前に、よくある `vda_model` のスペルミスを拒否し、正しい `vad_model` を案内します。これにより、VAD に依存する分割、話者処理、`sentence_info` が気付かないまま無効になることを防ぎます。GitHub のソースリリースでは legacy WebSocket ファイルランタイムも更新され、クライアントは明示的な入力終了応答を待ち、サーバーは保留中の offline、online、2pass 音声を処理してから完了またはエラーを返します。Python パッケージのインストール：`python -m pip install -U "funasr==1.4.0"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.0)
-- 2026/07/27：**v1.3.30 を PyPI に公開** — WAV、MP3、FLAC、OGG、MP4/M4A、WebM などのコンテナ形式の音声バイト列を raw PCM と誤認せず、対応するコーデックでデコードするようになりました。OpenAI 互換レスポンスは話者ラベルを保持し、句読点が一致しない場合も VAD 区間時刻を保持します。信頼済みブラウザクライアント向け CORS と、vLLM の 30 秒 VAD 上限にも対応しました。GitHub Release には、デスクトップ／サーバー向け 9 ターゲットの最新 llama.cpp ビルドも同梱しています。インストール：`python -m pip install -U "funasr==1.3.30"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.30)
-- 2026/07/24：**v1.3.29 hotfix が PyPI に公開** — SenseVoice の長時間音声推論で word timestamp と句読点モデルがない場合でも、各 VAD 音声区間を `sentence_info` で返すようになりました。字幕クライアントは、ゼロ長またはメディア全体を覆う一つの cue ではなく、認識テキストと実際のミリ秒単位の開始・終了時刻を取得できます。インストール：`python -m pip install -U "funasr==1.3.29"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.29)
-- 2026/07/24：**v1.3.28 hotfix が PyPI に公開** — VAD で確定した realtime WebSocket の最終結果が短い接頭辞、反復 hallucination、または decode 例外へ退化した場合、現在の音声区間を連続して完全に覆う clean partial を保持します。短い STOP tail、VAD finalize、話者処理も同じ信頼できる完了経路に統一しました。SenseVoice 字幕は rich tag、句読点、word/BPE timestamp を正しく整列し、中国語を一つの cue に潰さず、英語表記も保持します。インストール：`python -m pip install -U "funasr==1.3.28"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.28)
-- 2026/07/24：**v1.3.27 が PyPI に公開** — OpenAI 互換サーバーが `verbose_json` で SenseVoice の検出言語を返し、vLLM fallback 後はキャッシュ済み Fun-ASR-Nano `AutoModel` を再利用します。vLLM/VAD 初期化と fallback の両方が失敗した場合は、半初期化状態を残さず、後続リクエストで再試行できます。インストール：`python -m pip install -U "funasr==1.3.27"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.27)
-- 2026/07/23：**v1.3.26 が PyPI に公開** — `funasr-server --model fun-asr-nano --hub ms` は ModelScope hub 指定を vLLM 経路と AutoModel fallback の両方で尊重します。インストール：`python -m pip install -U "funasr==1.3.26"`。[Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.26)
-- 2026/07/23：**llama.cpp runtime v0.1.9** — Windows Vulkan 向け `funasr-llamacpp-windows-x64-vulkan.zip` を追加。現在の AMD、Intel、NVIDIA Vulkan ドライバーで SenseVoiceSmall を単独実行できます。Linux Vulkan、Windows CUDA、CPU/AVX2、Linux arm64、macOS arm64 も引き続き提供します。[Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.9)
-- 2026/07/22：**llama.cpp runtime v0.1.8** — Linux Vulkan tarball と Windows CUDA zip を含む CPU/エッジ向け GGUF ランタイム。現在のダウンロードと手順：[funasr.com/deploy/llama-cpp](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.8)
-- 2026/05/24：**v1.3.3** — `funasr-server` CLI、OpenAI互換API、MCPサーバー。`pip install --upgrade funasr`
-- 2026/05/20：Qwen3-ASR (0.6B/1.7B) 追加 — 52言語対応。
-- 2026/05/20：GLM-ASR-Nano (1.5B) 追加 — 17言語、方言対応。
-- 2026/05/19：Fun-ASR-NanoとSenseVoiceは、VADおよびCAM++と組み合わせて話者分離パイプラインを構成できます。
-- 2025/12/15：[Fun-ASR-Nano-2512](https://github.com/QwenAudio/Fun-ASR) — 中・英・日と中国語方言に対応。
+- **FunASR 1.4.11** は現在の PyPI 安定版です。実際のモデル timestamp を保持しながら、多言語字幕の可読性を改善しました。更新：`python -m pip install -U "funasr==1.4.11"`。[Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.11)
+- **MOSS-Transcribe-Diarize 連携**では、長時間音声の ASR、timestamp、匿名 speaker label を 1 回の生成で処理でき、外部 VAD や speaker pipeline は不要です。FunASR は local Transformers と既存の vLLM/SGLang service に対応し、FunClip は speaker 付き SRT と clip を出力できます。[Deployment guide ->](./docs/moss_transcribe_diarize.md)
+- **llama.cpp runtime v0.2.6** は Linux、macOS、Windows の 10 target 向け検証済み archive を提供し、RTX 30/40 と RTX 50 向け Windows CUDA package を分けて用意しています。[Download matrix ->](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6)
+- **Realtime serving の高速化と安定化**：互換 WebSocket session を直列処理せず batch 化し、decode queue によって正常な session を default で切断しないようにしました。H100 regression workload では 12-client STOP p95 が 19.8 秒から 0.4 秒に短縮しました。[Production guide ->](./docs/vllm_guide.md)
+
+> 完全な変更履歴と download asset は [GitHub Releases](https://github.com/modelscope/FunASR/releases) を参照してください。
 
 ---
 
diff --git a/README_ko.md b/README_ko.md
index 535fe394f..192142ef5 100644
--- a/README_ko.md
+++ b/README_ko.md
@@ -99,36 +99,12 @@ Whisper는 단일 모델이지만, **FunASR는 툴킷**입니다. 용도에 맞
 
 ## 최신 소식
 
-- 2026/08/31: **v1.4.11 PyPI 공개** — timestamp가 있는 ASR word에 SentencePiece `▁` boundary marker가 포함될 때 다국어 readable subtitle의 punctuation이 사라지는 문제를 수정했습니다. 이전에는 보이지 않는 marker 하나가 전체 녹음의 punctuation-to-timestamp alignment를 실패시켜 top-level transcript에 punctuation이 있어도 punctuation-free VAD segment로 fallback했습니다. 제보자의 10분 33초 sample에서는 예측된 punctuation 272개가 275개 timestamp sentence와 137개 readable cue에 모두 유지되며, 최대 7.94초·42자입니다. 정상 comparison sample의 alignment도 그대로 유지됩니다. 업데이트: `python -m pip install -U "funasr==1.4.11"`. [#3539](https://github.com/modelscope/FunASR/issues/3539)는 reporter 확인까지 open 상태로 유지합니다. [Fix ->](https://github.com/modelscope/FunASR/pull/3587) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.11)
-- 2026/08/31: **v1.4.10 PyPI 공개** — 너무 긴 subtitle을 나눌 때 중국어 단어 경계를 깨뜨리는 문제를 수정했습니다. Fallback은 글자 수 제한까지 기계적으로 채우지 않고 8초·42자 hard limit 안에서 duration과 length의 균형을 맞추며 Jieba lexical boundary, punctuation, script transition, 실제 timestamp gap을 우선합니다. 제보자와 동일한 10분 33초 중국어·한국어 mixed sample에서 84 source segments가 133 cues가 되었고 모든 cue가 제한 안에 있으며 `钟书成为`, `扭了一下` 같은 phrase도 유지됩니다. 업데이트: `python -m pip install -U "funasr==1.4.10"`. [#3539](https://github.com/modelscope/FunASR/issues/3539)는 reporter 확인까지 open 상태로 유지합니다. [Fix ->](https://github.com/modelscope/FunASR/pull/3583) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.10)
-- 2026/08/30: **v1.4.9 PyPI 공개** — v1.4.8 공개 후 merge된 subtitle 수정을 바로 설치할 수 있게 배포합니다. Readable SRT는 aligned word timestamp를 source sentence에 유지하고, 이미 너무 긴 sentence를 model의 실제 timestamp boundary에서만 나눕니다. 제보자의 10분 33초 중국어·한국어 mixed sample에서는 84 cues(8초 초과 26개)가 131 cues로 바뀌었고, 8초 또는 42자를 넘는 cue, overlap, fabricated timestamp는 0개였습니다. 업데이트: `python -m pip install -U "funasr==1.4.9"`. [#3539](https://github.com/modelscope/FunASR/issues/3539)는 reporter 재검증까지 open 상태로 유지합니다. [Fix ->](https://github.com/modelscope/FunASR/pull/3574) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.9)
-- 2026/08/30: **v1.4.8 PyPI 공개** — v1.4.7 이후 merge된 third-party MOSS-Transcribe-Diarize vLLM path를 완성했습니다. `backend="vllm"`은 공식 `response_format=diarized_json` speaker segment를 받아 FunASR `sentence_info`로 정규화합니다. 긴 회의에서는 `max_completion_tokens`를 server에 전달하며, default 5120-token 경계를 넘는 녹음에는 8192 tokens를 안내합니다. MOSS는 transcription, timestamp, speaker diarization을 end-to-end로 수행하므로 외부 `vad_model`이나 `spk_model`을 추가하지 않습니다. 업데이트: `python -m pip install -U "funasr==1.4.8"`. GitHub Release는 검증된 llama.cpp v0.2.6 10개 platform archive와 함께 제공합니다. [MOSS guide ->](./docs/moss_transcribe_diarize.md) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.8)
-- 2026/08/30: **llama.cpp runtime v0.2.6** — RTX 50 / Blackwell용 Windows CUDA architecture 120(`sm_120`) 전용 archive를 추가하면서 architecture 86 package도 유지합니다. 두 CUDA ZIP은 필요한 NVIDIA cuBLAS DLL과 license를 포함하고 MSVC runtime을 static link하며 PE import audit를 통과합니다. Linux, macOS, Windows용 archive 10개는 하나의 exact release commit에서 build하고 공개 SHA-256과 대조합니다. Build 및 package 검증은 Blackwell 실제 hardware inference 성공을 증명하지 않으므로 해당 archive로 재검증할 때까지 hardware report를 open 상태로 유지합니다. [구현 →](https://github.com/modelscope/FunASR/pull/3570) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6)
-- 2026/08/30: **v1.4.7 PyPI 공개** — OpenMOSS의 third-party model MOSS-Transcribe-Diarize를 FunASR `AutoModel`에 통합했습니다. local Transformers(`backend="hf"`) 또는 기존 vLLM server(`backend="vllm"`)를 선택할 수 있으며, 두 경로 모두 tag가 포함된 model 출력을 `text`, `raw_text`, millisecond `timestamp`, speaker-aware `sentence_info`로 정규화합니다. MOSS는 한 번의 inference로 transcription, timestamp, speaker diarization을 수행하므로 외부 `vad_model`과 `spk_model`을 지정하지 않습니다. 또한 동일 speaker의 매우 짧은 gap을 잇는 SRT cue 연속성을 개선하고 long session latency 진단용 opt-in realtime decode profiling을 추가했습니다. 업데이트: `python -m pip install -U "funasr==1.4.7"`. GitHub Release에는 검증된 llama.cpp v0.2.5 9개 platform runtime archive도 포함됩니다. [MOSS guide →](./docs/moss_transcribe_diarize.md) · [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.7)
-- 2026/08/29: **llama.cpp runtime v0.2.5** — graph 실행 전에 host weight를 선택한 Vulkan backend buffer에 upload합니다. Q8/F16 weight는 Linux Vulkan llvmpipe에서 검증했으며, Linux, macOS, Windows용 archive 9개를 exact commit `f371370d4c5e4c61d13d4eb9c55cda2f4dd95e4f`에서 build하고 공개 SHA-256과 대조했습니다. AMD Windows hardware crash가 수정됐다고 주장하지 않으며, [#3479](https://github.com/modelscope/FunASR/issues/3479)는 제보자의 실제 hardware 재검증을 위해 open 상태를 유지합니다. [수정 →](https://github.com/modelscope/FunASR/pull/3555) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.5)
-- 2026/08/29: **llama.cpp runtime v0.2.4** — F16 GGUF model에서 transcript가 간헐적으로 비는 문제를 수정했습니다. Query embedding은 F16 storage를 F32로 읽지 않고 GGML F16/F32 type에 맞게 decode됩니다. 정확한 v0.2.3 AVX2 asset은 완료된 298회 실행 중 22회의 빈 출력을 재현했지만, 수정 후에는 100/100회 동일한 비어 있지 않은 결과를 냈고 Q8 model과 byte 단위로 일치했습니다. Release workflow가 동일한 exact commit에서 9개 platform archive를 빌드합니다. [수정 →](https://github.com/modelscope/FunASR/pull/3550) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.4)
-- 2026/08/29: **v1.4.6 PyPI 공개** — realtime WebSocket server는 20초 간격 ping을 유지하면서 decode queue 때문에 pong 처리가 늦어진 정상 session을 기본적으로 종료하지 않습니다. 운영 환경의 queue 및 decode latency를 측정한 뒤 양수 timeout을 명시적으로 설정할 수도 있습니다. 빈 영어 timestamp segment에서 발생하는 `IndexError`를 방지하고 인접 subtitle word를 읽기 쉬운 cue로 묶습니다. 업데이트: `python -m pip install -U "funasr==1.4.6"`. GitHub Release에는 검증된 llama.cpp v0.2.3 9개 platform runtime archive도 포함됩니다. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.6)
-- 2026/08/29: **llama.cpp runtime v0.2.3** — backend 초기화 이후 model load, audio/VAD, graph build와 allocation, compute 단계마다 즉시 flush되는 경계 로그를 추가했습니다. `vulkan backend ready` 이후 발생하는 Windows AMD Vulkan `0xC0000005`의 위치를 좁히기 위한 진단이며, hardware별 crash가 수정됐다고 주장하지 않습니다. 동일한 exact commit에서 Linux, macOS, Windows용 archive 9개를 공개합니다. [문제 해결 경계 →](./runtime/llama.cpp/README.md#optional-windows-vulkan-backend-for-sensevoicesmall) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.3)
-- 2026/08/28: **v1.4.5 PyPI 공개** — `torchaudio`는 더 이상 inference의 필수 dependency가 아닙니다. Feature extraction은 사용할 수 있으면 `torchaudio.compliance.kaldi`를 사용하고, 그렇지 않으면 선택형 `kaldi-native-fbank` backend를 사용할 수 있습니다. 여전히 `torchaudio`가 필요한 작업은 해결 방법이 포함된 dependency 메시지를 반환합니다. 이 fallback은 Ascend 910B에서 end-to-end로 검증됐으며 70.47초 오디오를 1.15초(RTF 0.016)에 처리했습니다. 업데이트: `python -m pip install -U "funasr==1.4.5"`. `torchaudio`가 없는 환경: `python -m pip install -U "funasr[knf]==1.4.5"`. GitHub Release에는 검증된 llama.cpp v0.2.1 9개 platform runtime archive도 함께 제공됩니다. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.5)
-- 2026/08/27: **llama.cpp runtime v0.2.1** — Vulkan device 선택이 일치하는 integrated GPU를 허용하며, 일치하는 discrete GPU가 함께 있으면 이를 우선하고 없으면 iGPU로 fallback합니다. SHA-256을 공개한 Linux, macOS, Windows용 archive 9개를 다시 빌드하고 검증했습니다. Radeon 780M은 제보자의 실제 하드웨어 확인이 필요하며, 별도 RX 9070 XT `0xC0000005` 초기화 crash는 수정됐다고 주장하지 않습니다. [다운로드 목록 및 quickstart →](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.1)
-- 2026/08/26: **v1.4.4 PyPI 공개** — realtime WebSocket decode가 모든 connection을 하나의 process-wide lock으로 직렬화하지 않고 호환되는 session을 batch 처리합니다. H100 regression workload에서 12-client STOP p95는 19.8초에서 0.4초로 줄었고, 16-client aggregate throughput은 8.6x에서 13.2x로 향상됐으며 client error는 0건이었습니다. 이 hotfix는 `torch.amp`가 없는 PyTorch release와의 호환성도 복원하고 runtime binding에서 실제 예외를 발생시킵니다. 업데이트: `python -m pip install -U "funasr==1.4.4"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.4)
-- 2026/08/21: **v1.4.3 PyPI 공개** — `AutoModel(vad_model="silero-vad")`에서 선택형 Silero VAD adapter를 사용할 수 있으며, 밀리초 단위 구간, threshold, 8/16 kHz 입력, ONNX mode, 최대 구간 길이를 지원합니다. 업데이트: `python -m pip install -U "funasr==1.4.3"`. Adapter 활성화: `python -m pip install -U "funasr[silero]==1.4.3"`. 화자 수를 알고 있는 speaker diarization은 대규모 embedding을 메모리 사용량이 큰 dense spectral clustering 대신 fixed-K clustering으로 처리합니다. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.3)
-- 2026/08/14: **v1.4.2 PyPI 공개** — 문장부호 모델의 token 경계가 타임스탬프가 있는 ASR 단어 내부에 놓여도 문장 정렬이 올바른 자막 분할을 유지합니다. 분산 학습은 각 gradient accumulation window의 마지막 microbatch에서 DDP/FSDP gradient를 동기화하고, 해석된 설정에서 DeepSpeed/FSDP mode를 초기화합니다. 해당 GitHub 소스 tag에는 llama.cpp SRT 출력과 v0.2.0 AMD Vulkan submission 업데이트도 포함됩니다. 설치: `python -m pip install -U "funasr==1.4.2"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.2)
-- 2026/08/11: **llama.cpp runtime v0.2.0** — upstream llama.cpp를 `803b7fca`로 고정하고, 하나의 검증된 workflow에서 SHA-256 값이 제공되는 Linux, macOS, Windows용 archive 9개를 공개했습니다. Fun-ASR-Nano, SenseVoice 및 Paraformer CLI는 SRT 자막을 출력할 수 있으며, Vulkan 시작 시 AMD 진단 정보와 CPU fallback을 안내합니다. AMD Windows Vulkan crash 수정은 제보자의 실제 하드웨어 확인을 기다리고 있습니다. [다운로드 목록 및 quickstart →](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.0)
-- 2026/08/04: **v1.4.1 PyPI 공개** — Hugging Face의 `paraformer-en` 별칭이 중국어 모델을 잘못 내려받지 않고 공식 영어 checkpoint를 가리키도록 수정했습니다. 이 패치에는 Fun-ASR-Nano LoRA 미세 조정과 더 안전한 checkpoint 처리도 포함됩니다. 해당 GitHub 소스 tag에는 JSONL 타임스탬프 출력, SenseVoice TensorRT 배포, OpenClaw 실시간 전사 연동도 포함됩니다. 설치: `python -m pip install -U "funasr==1.4.1"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.1)
-- 2026/07/31: **v1.4.0 PyPI 공개** — `AutoModel`은 모델 다운로드 전에 흔한 `vda_model` 오타를 거부하고 올바른 `vad_model` 인수를 안내합니다. 따라서 VAD 기반 분할, 화자 처리 및 `sentence_info`가 조용히 비활성화되지 않습니다. GitHub 소스 릴리스에서는 legacy WebSocket 파일 런타임도 개선되어, 클라이언트가 명시적인 입력 종료 응답을 기다리고 서버가 대기 중인 offline, online 및 2pass 오디오를 처리한 뒤 완료 또는 오류를 반환합니다. Python 패키지 설치: `python -m pip install -U "funasr==1.4.0"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.4.0)
-- 2026/07/27: **v1.3.30 PyPI 공개** — WAV, MP3, FLAC, OGG, MP4/M4A, WebM 컨테이너 형식의 오디오 바이트를 raw PCM으로 잘못 해석하지 않고 해당 코덱으로 디코딩합니다. OpenAI 호환 응답은 화자 라벨을 보존하고, 구두점이 일치하지 않아도 VAD 문장 시간을 유지합니다. 신뢰할 수 있는 브라우저 클라이언트용 CORS와 vLLM의 30초 VAD 구간 제한도 지원합니다. GitHub Release에는 데스크톱 및 서버용 9개 대상의 최신 llama.cpp 빌드도 함께 제공합니다. 설치: `python -m pip install -U "funasr==1.3.30"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.30)
-- 2026/07/24: **v1.3.29 hotfix PyPI 공개** — SenseVoice 장시간 오디오 추론에서 word timestamp와 구두점 모델이 없을 때도 각 VAD 음성 구간을 `sentence_info`로 반환합니다. 자막 클라이언트는 길이가 0이거나 미디어 전체를 덮는 단일 cue 대신 인식 텍스트와 실제 밀리초 단위 시작·종료 시간을 받을 수 있습니다. 설치: `python -m pip install -U "funasr==1.3.29"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.29)
-- 2026/07/24: **v1.3.28 hotfix PyPI 공개** — VAD로 확정된 realtime WebSocket 최종 결과가 짧은 접두사, 반복 hallucination 또는 decode 예외로 퇴화하면 현재 음성 구간을 연속해서 완전히 덮는 clean partial을 보존합니다. 짧은 STOP tail, VAD finalize, 화자 완료 처리도 동일한 안정적인 경로로 통합했습니다. SenseVoice 자막은 rich tag, 구두점, word/BPE timestamp를 올바르게 정렬해 중국어가 하나의 cue로 합쳐지거나 영어 원문이 손상되지 않습니다. 설치: `python -m pip install -U "funasr==1.3.28"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.28)
-- 2026/07/24: **v1.3.27 PyPI 공개** — OpenAI 호환 서버가 `verbose_json`에 SenseVoice 감지 언어를 반환하고, vLLM fallback 후에는 캐시된 Fun-ASR-Nano `AutoModel`을 재사용합니다. vLLM/VAD 초기화와 fallback이 모두 실패하면 반쯤 초기화된 상태를 남기지 않아 이후 요청에서 다시 시도할 수 있습니다. 설치: `python -m pip install -U "funasr==1.3.27"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.27)
-- 2026/07/23: **v1.3.26 PyPI 공개** — `funasr-server --model fun-asr-nano --hub ms`가 vLLM 경로와 AutoModel fallback 모두에서 ModelScope hub 선택을 존중합니다. 설치: `python -m pip install -U "funasr==1.3.26"`. [Release →](https://github.com/modelscope/FunASR/releases/tag/v1.3.26)
-- 2026/07/23: **llama.cpp runtime v0.1.9** — Windows Vulkan용 `funasr-llamacpp-windows-x64-vulkan.zip`을 추가했습니다. 최신 AMD, Intel 또는 NVIDIA Vulkan 드라이버에서 SenseVoiceSmall을 독립 실행할 수 있습니다. Linux Vulkan, Windows CUDA, CPU/AVX2, Linux arm64, macOS arm64 패키지도 계속 제공합니다. [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.9)
-- 2026/07/22: **llama.cpp runtime v0.1.8** — Linux Vulkan tarball과 Windows CUDA zip을 포함한 CPU/엣지용 GGUF 런타임입니다. 현재 다운로드와 빠른 시작: [funasr.com/deploy/llama-cpp](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.8)
-- 2026/05/24: **v1.3.3** — `funasr-server` CLI, OpenAI 호환 API, MCP 서버. `pip install --upgrade funasr`
-- 2026/05/20: Qwen3-ASR (0.6B/1.7B) 추가 — 52개 언어 지원.
-- 2026/05/20: GLM-ASR-Nano (1.5B) 추가 — 17개 언어, 방언 지원.
-- 2026/05/19: Fun-ASR-Nano 및 SenseVoice는 VAD 및 CAM++와 결합하여 화자 분리 파이프라인을 구성할 수 있습니다.
-- 2025/12/15: [Fun-ASR-Nano-2512](https://github.com/QwenAudio/Fun-ASR) — 중국어, 영어, 일본어 및 중국어 방언 지원.
+- **FunASR 1.4.11**은 현재 PyPI 안정 버전입니다. 실제 모델 timestamp를 유지하면서 다국어 자막의 가독성을 개선했습니다. 업데이트: `python -m pip install -U "funasr==1.4.11"`. [Release ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.11)
+- **MOSS-Transcribe-Diarize 연동**은 긴 오디오의 ASR, timestamp, 익명 speaker label을 한 번의 생성으로 처리하므로 외부 VAD 또는 speaker pipeline이 필요하지 않습니다. FunASR은 local Transformers와 기존 vLLM/SGLang service를 지원하고, FunClip은 speaker별 SRT와 clip을 내보낼 수 있습니다. [Deployment guide ->](./docs/moss_transcribe_diarize.md)
+- **llama.cpp runtime v0.2.6**은 Linux, macOS, Windows 10개 target용 검증된 archive를 제공하며 RTX 30/40 및 RTX 50용 Windows CUDA package를 별도로 제공합니다. [Download matrix ->](https://www.funasr.com/en/deploy/llama-cpp.html) · [Release ->](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6)
+- **Realtime serving 성능과 안정성 개선**: 호환 WebSocket session을 직렬 처리하지 않고 batch 처리하며, decode queue 때문에 정상 session을 기본적으로 종료하지 않습니다. H100 regression workload에서 12-client STOP p95가 19.8초에서 0.4초로 줄었습니다. [Production guide ->](./docs/vllm_guide.md)
+
+> 전체 변경 기록과 download asset은 [GitHub Releases](https://github.com/modelscope/FunASR/releases)에서 확인할 수 있습니다.
 
 ---
 
diff --git a/README_zh.md b/README_zh.md
index 967519fdc..cfbc545b7 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -149,56 +149,12 @@ Whisper 是单个模型，**FunASR 是一个工具箱**——按场景挑模型
 
 ## 最新动态
 
-- 2026/08/31：**v1.4.11 已发布到 PyPI** — 修复中韩等多语言可读字幕因时间戳词元携带 SentencePiece `▁` 边界标记而丢失标点的问题。此前一个不可见标记就可能让整段标点与时间戳对齐失败，回退为无标点 VAD 片段，即使顶层识别文本已有标点。报告者的 10 分 33 秒样本现在将 272 个预测标点完整保留到 275 个时间戳句子和 137 条可读字幕中，最长 7.94 秒、最长 42 字；正常对照样本保持对齐。升级命令：`python -m pip install -U "funasr==1.4.11"`；[#3539](https://github.com/modelscope/FunASR/issues/3539) 将继续保持开放，等待报告者确认。[修复 ->](https://github.com/modelscope/FunASR/pull/3587) · [发布页 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.11)
-- 2026/08/31：**v1.4.10 已发布到 PyPI** — 修复超长字幕拆分时破坏中文词语边界的问题。回退算法不再机械填满字符上限，而是在 8 秒和 42 字符的硬限制内平衡时长与长度，并优先选择 Jieba 词边界、标点、文字系统切换和真实时间戳间隙。在报告者同一份 10 分 33 秒中韩混合样本上，84 个源片段被拆为 133 条字幕，全部满足限制，同时保留“钟书成为”“扭了一下”等完整短语。升级命令：`python -m pip install -U "funasr==1.4.10"`；[#3539](https://github.com/modelscope/FunASR/issues/3539) 将继续保持开放，等待报告者确认。[修复 ->](https://github.com/modelscope/FunASR/pull/3583) · [发布页 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.10)
-- 2026/08/30：**v1.4.9 已发布到 PyPI** — 将 v1.4.8 发布后合入的字幕修复交付为可直接安装的版本。可读 SRT 渲染现在会把对齐后的词级时间戳带入源句，并且只在模型的真实时间边界上拆分原本已经过长的句子。在报告者提供的 10 分 33 秒中韩混合样本上，精确合并代码把 84 段、其中 26 段超过 8 秒的字幕变为 131 段；最长不超过 8 秒或 42 个字符，无重叠，也不伪造时间戳。升级命令：`python -m pip install -U "funasr==1.4.9"`；[#3539](https://github.com/modelscope/FunASR/issues/3539) 会保持开放，等待报告者复测。[修复 ->](https://github.com/modelscope/FunASR/pull/3574) · [发布页 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.9)
-- 2026/08/30：**v1.4.8 已发布到 PyPI** — 补齐 v1.4.7 之后合入的第三方 MOSS-Transcribe-Diarize vLLM 能力。`backend="vllm"` 现在可接收官方 `response_format=diarized_json` 说话人分段，并统一为 FunASR `sentence_info`；长会议可通过 `max_completion_tokens` 传递生成上限，文档对超过默认 5120-token 边界的录音给出 8192-token 配置。MOSS 仍然端到端完成转写、时间戳和说话人识别，不要外挂 `vad_model` 或 `spk_model`。升级命令：`python -m pip install -U "funasr==1.4.8"`。GitHub 发布页同时提供已验证的 llama.cpp v0.2.6 十平台运行包。[MOSS 部署指南 ->](./docs/moss_transcribe_diarize_zh.md) · [发布页 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.8)
-- 2026/08/30：**llama.cpp runtime v0.2.6** — 新增面向 RTX 50 / Blackwell 的 Windows CUDA architecture 120（`sm_120`）专用包，同时保留 architecture 86 包。两个 CUDA ZIP 都包含所需的 NVIDIA cuBLAS DLL 与许可证，静态链接 MSVC runtime，并通过 PE 导入审计；十个 Linux、macOS 与 Windows 压缩包由同一个准确发布提交构建，并按公开 SHA-256 逐一复核。构建与打包通过不代表 Blackwell 实机推理已经验证，因此硬件报告会保持开放，等待用户使用匹配资产复测。[实现 →](https://github.com/modelscope/FunASR/pull/3570) · [发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.6)
-- 2026/08/30：**v1.4.7 已发布到 PyPI** — OpenMOSS 的第三方模型 MOSS-Transcribe-Diarize 已接入 FunASR `AutoModel`。可选择本地 Transformers（`backend="hf"`）或已有 vLLM 服务（`backend="vllm"`）；两条路径都会将模型标签输出统一为 `text`、`raw_text`、毫秒级 `timestamp` 和带说话人标签的 `sentence_info`。MOSS 在一次推理中完成转写、时间戳和说话人分离，因此不要再外挂 `vad_model` 或 `spk_model`。本版本还改善了同一说话人极短间隔处的 SRT 字幕连续性，并增加可选的实时解码性能统计，便于定位长会话延迟。升级命令：`python -m pip install -U "funasr==1.4.7"`。GitHub 发布页同时提供已验证的 llama.cpp v0.2.5 九平台运行包。[MOSS 部署指南 →](./docs/moss_transcribe_diarize_zh.md) · [发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.4.7)
-- 2026/08/29：**llama.cpp runtime v0.2.5** — 在计算图执行前，将 host 权重上传到所选 Vulkan backend buffer。Q8 与 F16 权重已通过本地 Linux Vulkan llvmpipe 验证，九个 Linux、macOS 与 Windows 压缩包均由准确提交 `f371370d4c5e4c61d13d4eb9c55cda2f4dd95e4f` 构建，并按公开 SHA-256 逐一复核。本版本不宣称 AMD Windows 硬件崩溃已经修复；[#3479](https://github.com/modelscope/FunASR/issues/3479) 保持开放，等待报告者实机复测。[修复 →](https://github.com/modelscope/FunASR/pull/3555) · [发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.5)
-- 2026/08/29：**llama.cpp runtime v0.2.4** — 修复 F16 GGUF 模型偶发空转写。运行时现在按 GGML 的 F16/F32 类型解码查询 embedding，不再把 F16 存储误读为 F32。精确的 v0.2.3 AVX2 发布包在 298 次完整运行中复现 22 次空结果；修复后 100/100 次输出稳定唯一，并与 Q8 模型逐字节一致。发布工作流将在同一准确提交上构建九个 Linux、macOS 与 Windows 压缩包。[修复 →](https://github.com/modelscope/FunASR/pull/3550) · [发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.4)
-- 2026/08/29：**v1.4.6 已发布到 PyPI** — 实时 WebSocket 服务继续每 20 秒发送 ping，但当排队解码延迟 pong 处理时，默认不再误关仍健康的连接；运维人员仍可在测量生产队列与解码延迟后显式配置正数超时。本版本同时避免空英文时间戳片段触发 `IndexError`，并将相邻字幕词组合为更易读的字幕段。升级命令：`python -m pip install -U "funasr==1.4.6"`。GitHub 发布页同时提供已验证的 llama.cpp v0.2.3 九平台运行包。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.4.6)
-- 2026/08/29：**llama.cpp runtime v0.2.3** — 后端初始化完成后新增可立即刷新的阶段边界，覆盖模型加载、音频/VAD、计算图构建与分配、推理计算，可继续缩小发生在 `vulkan backend ready` 之后的 Windows AMD Vulkan `0xC0000005` 崩溃范围；本版本不宣称已经修复该硬件相关崩溃。九个 Linux、macOS 与 Windows 压缩包由同一准确提交构建并发布。[排障边界 →](./runtime/llama.cpp/README.md#optional-windows-vulkan-backend-for-sensevoicesmall) · [发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.3)
-- 2026/08/28：**v1.4.5 已发布到 PyPI** — `torchaudio` 不再是推理的硬依赖。特征提取会优先使用 `torchaudio.compliance.kaldi`，不可用时可切换到可选的 `kaldi-native-fbank` 后端；仍然必须使用 `torchaudio` 的操作会给出可执行的依赖提示。该 fallback 已在 Ascend 910B 上端到端验证：70.47 秒音频用时 1.15 秒（RTF 0.016）。常规升级：`python -m pip install -U "funasr==1.4.5"`；没有 `torchaudio` 的环境请安装：`python -m pip install -U "funasr[knf]==1.4.5"`。GitHub 发布页同时提供已验证的 llama.cpp v0.2.1 九平台运行包。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.4.5)
-- 2026/08/27：**llama.cpp runtime v0.2.1** — Vulkan 设备选择现在接受匹配的集成 GPU；同时存在匹配独显时优先独显，否则回退到核显。九个 Linux、macOS 与 Windows 压缩包已重新构建并通过公开 SHA-256 复核。Radeon 780M 仍需报告者实机确认；单独的 RX 9070 XT `0xC0000005` 初始化崩溃不在本次修复声明内。[下载矩阵与快速开始 →](https://www.funasr.com/deploy/llama-cpp.html) · [发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.1)
-- 2026/08/26：**v1.4.4 已发布到 PyPI** — 实时 WebSocket 解码会对兼容会话进行批处理，不再让所有连接排队经过同一个进程级锁。在 H100 回归负载下，12 路 STOP p95 从 19.8 秒降至 0.4 秒，16 路聚合吞吐从 8.6x 提升到 13.2x，且客户端零错误。本热修复同时兼容不提供 `torch.amp` 的 PyTorch 版本，并让运行时绑定抛出真实异常。升级命令：`python -m pip install -U "funasr==1.4.4"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.4.4)
-- 2026/08/21：**v1.4.3 已发布到 PyPI** — `AutoModel(vad_model="silero-vad")` 现可选用 Silero VAD 适配器，返回毫秒级片段，并支持阈值、8/16 kHz 输入、ONNX 模式和最长片段限制。基础升级命令：`python -m pip install -U "funasr==1.4.3"`；启用该适配器：`python -m pip install -U "funasr[silero]==1.4.3"`。已知说话人数的说话人分离在大规模 embedding 输入下改用固定 K 聚类，避免内存开销较高的稠密谱聚类。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.4.3)
-- 2026/08/14：**v1.4.2 已发布到 PyPI** — 标点模型的 token 边界落在带时间戳的 ASR 词内部时，句子对齐现在仍能保留正确的字幕分段。分布式训练会在每个梯度累积窗口的最后一个 microbatch 同步 DDP/FSDP 梯度，并从解析后的配置正确初始化 DeepSpeed/FSDP 模式。对应 GitHub 源码 tag 同时包含 llama.cpp SRT 输出和 v0.2.0 AMD Vulkan submission 更新。安装命令：`python -m pip install -U "funasr==1.4.2"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.4.2)
-- 2026/08/11：**llama.cpp runtime v0.2.0** — 统一固定上游 llama.cpp 到 `803b7fca`，通过同一套测试工作流发布 9 个带 SHA-256 校验值的 Linux、macOS 与 Windows 压缩包。Fun-ASR-Nano、SenseVoice 和 Paraformer CLI 现在可直接输出 SRT 字幕；Vulkan 启动会给出可操作的 AMD 诊断信息和 CPU fallback。AMD Windows Vulkan 崩溃修复仍等待 issue 报告者在原硬件上确认。[下载矩阵与快速开始 →](https://www.funasr.com/deploy/llama-cpp.html) · [发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.2.0)
-- 2026/08/04：**v1.4.1 已发布到 PyPI** — Hugging Face 的 `paraformer-en` 别名现在会解析到官方英文 checkpoint，不再静默下载中文模型。本补丁还包含 Fun-ASR-Nano LoRA 微调与更安全的 checkpoint 处理；对应 GitHub 源码 tag 同时提供 JSONL 时间戳输出、SenseVoice TensorRT 部署和 OpenClaw 实时转写集成。安装命令：`python -m pip install -U "funasr==1.4.1"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.4.1)
-- 2026/08/04：**OpenClaw 实时转写集成** — 新增 [`openclaw-funasr`](integrations/openclaw/) 源码包，把私有部署的 FunASR `online`、`offline` 与 `2pass` WebSocket 识别接入 OpenClaw Talk 和 Voice Call。8 kHz G.711 mu-law 转换、60 ms 分帧、partial/final 文本、重连上限、安装包与运行时注册均已基于 OpenClaw `2026.7.2` 验证；npm 与 ClawHub 发布将在所需的[上游 SDK 改动](https://github.com/openclaw/openclaw/pull/118977)合入后进行。
-- 2026/07/31：**v1.4.0 已发布到 PyPI** — `AutoModel` 现在会在下载模型前拒绝常见的 `vda_model` 误拼写并明确提示使用 `vad_model`，避免依赖 VAD 的分段、说话人处理和 `sentence_info` 被静默关闭。GitHub 源码发布同时更新 legacy WebSocket 文件运行时：客户端会等待明确的输入结束确认，服务端先刷新待处理的 offline、online 与 2pass 音频，并把收尾失败返回给客户端。Python 包安装命令：`python -m pip install -U "funasr==1.4.0"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.4.0)
-- 2026/07/27：**v1.3.30 已发布到 PyPI** — WAV、MP3、FLAC、OGG、MP4/M4A 和 WebM 等容器格式的音频字节现在会通过对应编解码器解码，不再被误当作原始 PCM。OpenAI 兼容响应会保留说话人标签，标点不匹配时仍保留 VAD 分句时间，受信任的浏览器客户端可按需启用 CORS，vLLM 的 VAD 分段上限为 30 秒。GitHub 发布页还同时提供覆盖九种桌面和服务器目标的当前 llama.cpp 预编译运行包。安装命令：`python -m pip install -U "funasr==1.3.30"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.30)
-- 2026/07/24：**v1.3.29 热修复已发布到 PyPI** — SenseVoice 长音频在没有词级时间戳和标点模型时，现在会通过 `sentence_info` 返回每个 VAD 语音片段。字幕客户端可直接获得识别文本及真实的毫秒级起止时间，不再退化为零时长或覆盖整段媒体的单条字幕。安装命令：`python -m pip install -U "funasr==1.3.29"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.29)
-- 2026/07/24：**v1.3.28 热修复已发布到 PyPI** — 实时 WebSocket 在 VAD 锁句结果退化为短前缀、重复幻觉或解码异常时，会保留连续且完整覆盖当前语音段的干净 partial；短音频 STOP、VAD 收尾和说话人结束现在统一走可靠的完成路径。SenseVoice 字幕分句也会正确对齐富标签、标点与词/BPE 时间戳，不再把中文压成一个字幕块，也不会破坏英文原文。安装命令：`python -m pip install -U "funasr==1.3.28"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.28)
-- 2026/07/24：**v1.3.27 已发布到 PyPI** — OpenAI 兼容服务现在会在 `verbose_json` 中返回 SenseVoice 检测到的语言，并在 vLLM 降级后复用已缓存的 Fun-ASR-Nano `AutoModel`。当 vLLM/VAD 初始化及其 fallback 均失败时，不会残留半初始化的 engine 状态，后续请求可以重试。安装命令：`python -m pip install -U "funasr==1.3.27"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.27)
-- 2026/07/23：**llama.cpp runtime v0.1.9** — 新增独立的 Windows Vulkan 包 `funasr-llamacpp-windows-x64-vulkan.zip`，支持在安装当前 AMD、Intel 或 NVIDIA Vulkan 驱动的 Windows 机器上运行 SenseVoiceSmall；Linux Vulkan、Windows CUDA、CPU/AVX2、Linux arm64 和 macOS arm64 包继续提供。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.9)
-- 2026/07/23：**v1.3.26 已发布到 PyPI** — `funasr-server --model fun-asr-nano --hub ms` 现在会在默认 Fun-ASR-Nano 的 vLLM 路径和 AutoModel fallback 路径中都尊重 ModelScope hub 选择，避免用户指定 ModelScope 时仍误走 Hugging Face 下载。安装命令：`python -m pip install -U "funasr==1.3.26"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.26)
-- 2026/07/23：**v1.3.25 已发布到 PyPI** — 实时 WebSocket 服务新增 `POSTPROCESS_HOTWORDS:错词=>正确词` 与 `--postprocess-hotword-file`，可在 final 文本阶段做确定性热词纠正，避免把固定错词修正误用成模型层 `HOTWORDS:` 解码偏置；源码目录下的实时服务入口也可直接运行。安装命令：`python -m pip install -U "funasr==1.3.25"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.25)
-- 2026/07/23：**v1.3.24 已发布到 PyPI** — OpenAI 兼容服务现在支持自定义模型路径和 hub 选择，llama.cpp/GGUF 文档补充 HTTP 转写 wrapper 与 Linux Vulkan 包，公开文档链接也已刷新，便于新用户顺利上手。安装命令：`python -m pip install -U "funasr==1.3.24"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.24)
-- 2026/07/19：**v1.3.22 已发布到 PyPI** — `funasr-server` 现在会为 SenseVoice/Paraformer fallback 的纯文本结果补齐 OpenAI 兼容 `verbose_json.segments`，避免字幕类客户端在 `text` 已有内容时仍拿到空 `segments` 数组。安装命令：`python -m pip install -U "funasr==1.3.22"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.22)
-- 2026/07/19：**v1.3.21 已发布到 PyPI** — 修复全新环境里先安装 `funasr`、尚未选择平台对应 PyTorch 版本时的首次导入阻塞。现在 `import funasr` 和 `funasr.__version__` 不再因为缺少 torch 失败；真正访问 `AutoModel` 时仍会要求安装 PyTorch，并给出明确安装提示。安装命令：`python -m pip install -U "funasr==1.3.21"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.21)
-- 2026/07/19：**v1.3.20 已发布到 PyPI** — PyPI 项目页和安装引导已同步到当前 FunASR 文档、社区集成列表，以及 Fun-ASR-Nano 部署路径中带引号的 `python -m pip install -U "funasr>=1.3.19"` 命令。本版本是文档/打包元数据同步，运行时代码与 v1.3.19 保持一致。安装命令：`python -m pip install -U "funasr==1.3.20"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.20)
-- 2026/07/19：**v1.3.19 已发布到 PyPI** — 实时 WebSocket 长会话排障文档已随包发布。启动服务时加上 `--enable-spk --log-session-stats-interval 30`，如果仍遇到断连或内存增长，请在 issue 中附上输出的 `Session stats:` 日志。安装命令：`python -m pip install -U "funasr==1.3.19"`。[长会话诊断 →](docs/vllm_guide_zh.md#长会话诊断) · [发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.19)
-- 2026/07/19：**v1.3.18 已发布到 PyPI** — CLI 的 SRT/TSV 字幕输出现在会请求句级时间戳，并在需要时加载标点模型；`funasr audio.wav --output-format srt --output-dir ./subs` 会输出分句字幕，不再退化成一个全文字幕块。安装命令：`python -m pip install -U "funasr==1.3.18"`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/v1.3.18)
-- 2026/07/18：**v1.3.16 已发布到 PyPI** — Fun-ASR-Nano 实时服务新增客户端分句模式。一个 WebSocket 会话可连续发送 PCM，并用 `COMMIT` 提交每个句子；无需加载服务端 VAD，短句可正常结束，多轮时间戳保持递增。执行 `pip install --upgrade funasr` 后，可用 `funasr-realtime-server --endpoint-mode client` 启动。[使用文档 →](examples/industrial_data_pretraining/fun_asr_nano/docs/realtime_demo.md)
-- 2026/07/22：**llama.cpp runtime v0.1.8** — 新增 Linux Vulkan 预编译包 `funasr-llamacpp-linux-x64-vulkan.tar.gz`，可在支持 Vulkan driver/ICD 的 Linux GPU 上运行 `llama-funasr-sensevoice ... --backend vulkan`；CPU、AVX2、macOS arm64、Windows CPU/AVX2、Windows CUDA 包继续保留。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.8)
-- 2026/07/18：**llama.cpp runtime v0.1.7** — 新增 SenseVoiceSmall 的 Windows CUDA 预编译包 `funasr-llamacpp-windows-x64-cuda.zip`，并保留 Linux / macOS / Windows CPU 包。下载 GGUF 模型后，可在支持的 NVIDIA GPU 上运行 `llama-funasr-sensevoice ... --backend cuda`。[发布页 →](https://github.com/modelscope/FunASR/releases/tag/runtime-llamacpp-v0.1.7)
-- 2026/05/24：**vLLM 推理引擎** — Fun-ASR-Nano 解码加速 2-3 倍。支持流式 WebSocket 服务（VAD + 说话人分离 + 热词）。[文档 →](docs/vllm_guide_zh.md) · [实时 WS 调优 →](docs/vllm_guide_zh.md#67-生产并发与多进程部署) · [API 稳定性清单 →](docs/vllm_guide_zh.md#生产-api-稳定性清单)
-- 2026/05/24：**动态 VAD** — 自适应静音阈值（默认开启），短句不切碎、长句自动切分。[详情 →](docs/vllm_guide_zh.md#7-动态-vad)
-- 2026/05/24：**v1.3.3** — `funasr-server` 命令行工具、OpenAI 兼容 API、MCP 服务。`pip install --upgrade funasr`
-- 2026/05/20：新增 Qwen3-ASR (0.6B/1.7B)，52 种语言自动检测。[使用方法](examples/industrial_data_pretraining/qwen3_asr)
-- 2026/05/20：新增 GLM-ASR-Nano (1.5B)，17 种语言，方言优化。[使用方法](examples/industrial_data_pretraining/glm_asr)
-- 2026/05/19：Fun-ASR-Nano 和 SenseVoice 可与 VAD、CAM++ 组合为说话人分离 pipeline。
-- 2025/12/15：[Fun-ASR-Nano-2512](https://github.com/QwenAudio/Fun-ASR) 上线，支持中/英/日及中文方言。
-
-<details><summary>更早</summary>
-
-- 2024/10/10：支持 Whisper-large-v3-turbo。
-- 2024/07/04：[SenseVoice](https://github.com/QwenAudio/SenseVoice) 发布。
-- 2024/01/30：FunASR 1.0 发布。
+- **FunASR 1.4.11** 是当前 PyPI 稳定版，重点改善多语言可读字幕，并保留模型真实时间戳。升级命令：`python -m pip install -U "funasr==1.4.11"`。[发布页 ->](https://github.com/modelscope/FunASR/releases/tag/v1.4.11)
+- **MOSS-Transcribe-Diarize 已接入生态**：一次生成同时完成长音频转写、时间戳和匿名说话人标签，无需外挂 VAD 或说话人流水线。FunASR 支持本地 Transformers 及已有 vLLM/SGLang 服务，FunClip 可导出带说话人的 SRT 和片段。[部署指南 ->](./docs/moss_transcribe_diarize_zh.md)
+- **llama.cpp runtime v0.2.6** 提供覆盖 Linux、macOS、Windows 十种目标的已验证预编译包，并为 RTX 30/40 与 RTX 50 显卡分别提供 Windows CUDA 包。[下载矩阵 ->](https://www.funasr.com/deploy/llama-cpp.html)
+- **实时服务更快、更稳定**：兼容的 WebSocket 会话改为批处理，排队解码也不再默认误关健康连接。在 H100 回归负载下，12 路客户端 STOP p95 从 19.8 秒降至 0.4 秒。[生产部署指南 ->](./docs/vllm_guide_zh.md)
 
-</details>
+> 完整改动记录和可下载资产请查看 [GitHub Releases](https://github.com/modelscope/FunASR/releases)。
 
 ---
 
diff --git a/tests/test_docs_funasr_install_commands.py b/tests/test_docs_funasr_install_commands.py
index 74dcfbbfd..e069cbd96 100644
--- a/tests/test_docs_funasr_install_commands.py
+++ b/tests/test_docs_funasr_install_commands.py
@@ -341,8 +341,8 @@ def test_top_level_readmes_surface_current_release_and_edge_runtime():
     }
 
     for name, text in readmes.items():
-        assert 'python -m pip install -U "funasr==1.4.9"' in text, name
-        assert "https://github.com/modelscope/FunASR/releases/tag/v1.4.9" in text, name
+        assert 'python -m pip install -U "funasr==1.4.11"' in text, name
+        assert "https://github.com/modelscope/FunASR/releases/tag/v1.4.11" in text, name
         assert "runtime-llamacpp-v0.2.6" in text, name
 
     assert "https://www.funasr.com/en/deploy/llama-cpp.html" in readmes["README.md"]
@@ -364,6 +364,21 @@ def test_top_level_readmes_surface_current_release_and_edge_runtime():
         assert "releases/download/runtime-llamacpp-v0.2.1/" not in text, name
 
 
+def test_top_level_readme_news_stays_concise():
+    headings = {
+        "README.md": "## What's new",
+        "README_zh.md": "## 最新动态",
+        "README_ja.md": "## 最新情報",
+        "README_ko.md": "## 최신 소식",
+    }
+
+    for name, heading in headings.items():
+        text = (ROOT / name).read_text()
+        news = text.split(heading, 1)[1].split("\n---", 1)[0]
+        assert news.count("\n- ") <= 5, name
+        assert "https://github.com/modelscope/FunASR/releases" in news, name
+
+
 def test_repository_roadmap_tracks_current_delivery_and_open_work():
     docs = [
         (ROOT / "docs/repository_roles.md").read_text(),