From 4b57640665ce2758923d5ada66bbf91984ce2901 Mon Sep 17 00:00:00 2001 From: Venkatesh Chandran Date: Mon, 10 Aug 2026 18:40:46 +0100 Subject: [PATCH 1/4] Document all-languages GPU image and SM_LANGUAGES for Container 15.18.0 - Add an All languages option for sm-gpu-inference-server-all-lang under the Enhanced model section, and note it in the section intro - Document SM_LANGUAGES on the GPU inference server, including the unset default and the startup failure on an unrecognised code - Bump latestContainerVersion to 15.18.0 - Record English-only language configuration in both benchmark tables and bump the batch benchmark version Co-Authored-By: Claude Opus 5 (1M context) --- .../deployments/container/accessing-images.mdx | 12 +++++++++++- .../container/gpu-speech-to-text.mdx | 12 ++++++++++++ .../container/performance-and-cost.mdx | 18 ++++++++++-------- sm-variables.ts | 2 +- 4 files changed, 34 insertions(+), 10 deletions(-) diff --git a/docs/deployments/container/accessing-images.mdx b/docs/deployments/container/accessing-images.mdx index 7dc5c2aa..3a793aed 100644 --- a/docs/deployments/container/accessing-images.mdx +++ b/docs/deployments/container/accessing-images.mdx @@ -78,7 +78,7 @@ docker pull speechmaticspublic.azurecr.io/sm-gpu-inference-server-en:${smVariabl ### Enhanced model -Depending on which Enhanced model languages are required, you can pull specific images. +Depending on which Enhanced model languages are required, you can pull specific images, or a single image covering all languages.
Language Pack 1 @@ -112,6 +112,16 @@ Depending on which Enhanced model languages are required, you can pull specific
+
+ All languages +

All Enhanced and Standard model languages

+ + {`docker pull speechmaticspublic.azurecr.io/sm-gpu-inference-server-all-lang:${smVariables.latestContainerVersion} `} + +
+ +To load only some of an image's languages at startup, see [Loading only selected languages](/deployments/container/gpu-speech-to-text#loading-only-selected-languages). + See [how to run the Transcription GPU container here.](/deployments/container/gpu-speech-to-text) :::info diff --git a/docs/deployments/container/gpu-speech-to-text.mdx b/docs/deployments/container/gpu-speech-to-text.mdx index fc82ca0a..de99cf78 100644 --- a/docs/deployments/container/gpu-speech-to-text.mdx +++ b/docs/deployments/container/gpu-speech-to-text.mdx @@ -117,6 +117,18 @@ To save GPU memory for throughput, you can run the server with only one model lo When running the all language standard model GPU inference server you must set the `SM_MODEL` environment variable to `standard` ::: +### Loading only selected languages + +Images that contain more than one language model, such as the [all language images](/deployments/container/accessing-images#enhanced-model), load every language at startup. To save GPU memory for throughput, pass the `SM_LANGUAGES` environment variable to the container and set it to a comma-separated list of language codes. Only those languages are loaded: + +```bash +-e SM_LANGUAGES=en,de,fr +``` + +Codes must be languages the image contains. A code the image does not contain fails startup with an error. + +When `SM_LANGUAGES` is unset, all of the image's languages are loaded. `SM_LANGUAGES` is available from Container version 15.18.0. + ### Monitoring the server The inference server is based on [Nvidia's Triton architecture](https://developer.nvidia.com/nvidia-triton-inference-server) and as such diff --git a/docs/deployments/container/performance-and-cost.mdx b/docs/deployments/container/performance-and-cost.mdx index b158c220..7e9392f9 100644 --- a/docs/deployments/container/performance-and-cost.mdx +++ b/docs/deployments/container/performance-and-cost.mdx @@ -24,14 +24,15 @@ This is a comparison of the performance and estimated running costs of transcrip The benchmark uses the following configuration: -| Benchmark details | | -|-------------------|----------------------------------------------| -| Version | 15.7.0 (Standard, Enhanced), 1.3.0 (Melia-1) | -| CPU | D16ds_v5 | -| GPU Standard | Standard_NC16as_T4_v3 | -| GPU Enhanced | Standard_NC8as_T4_v3 | -| GPU Melia-1 | Standard_NC40ads_H100_v5 | -| Price Basis | Azure PAYG East US, Linux, Standard | +| Benchmark details | | +|-------------------|-----------------------------------------------| +| Version | 15.18.0 (Standard, Enhanced), 1.3.0 (Melia-1) | +| Language | English only | +| CPU | D16ds_v5 | +| GPU Standard | Standard_NC16as_T4_v3 | +| GPU Enhanced | Standard_NC8as_T4_v3 | +| GPU Melia-1 | Standard_NC40ads_H100_v5 | +| Price Basis | Azure PAYG East US, Linux, Standard | :::note For GPU Models, transcribers and inference servers were all run on a single VM node. @@ -54,6 +55,7 @@ This benchmark uses the following configuration[^4]: | Benchmark details | Value | | ------------------| ----------------------------------- | | Version | 13.4.0 | +| Language | English only | | CPU | D16ds_v5 | | GPU Standard | Standard_NC16as_T4_v3 | | GPU Enhanced | Standard_NC8as_T4_v3 | diff --git a/sm-variables.ts b/sm-variables.ts index b6af3db1..1e520837 100644 --- a/sm-variables.ts +++ b/sm-variables.ts @@ -1,7 +1,7 @@ export const smVariables = { jsonOutputVersion: "2.9", latestApplianceVersion: "6.3.0", - latestContainerVersion: "15.14.0", + latestContainerVersion: "15.18.0", latestMelia1ContainerVersion: "1.3.0", usageContainerVersion: "0.3.0", helmChartVersion: "1.4.0", From effc8339018d0de76deb3f3127192e2459b4f6c0 Mon Sep 17 00:00:00 2001 From: Venkatesh Chandran Date: Tue, 11 Aug 2026 10:21:07 +0100 Subject: [PATCH 2/4] Move all-languages image to its own section and cover SM_MODEL alongside SM_LANGUAGES - Promote the all-languages image from a collapsible under Enhanced model to a Standard and Enhanced model section, and restore the Enhanced model intro - Hoist the run link and Support admonition to the transcription GPU intro so they scope to every image in the section rather than one subsection - Link the new section to both Running only one model and Loading only selected languages - Note that SM_LANGUAGES and SM_MODEL can be combined, and drop the single-image reference from the SM_LANGUAGES intro Co-Authored-By: Claude Opus 5 (1M context) --- .../container/accessing-images.mdx | 27 ++++++++++--------- .../container/gpu-speech-to-text.mdx | 4 ++- 2 files changed, 17 insertions(+), 14 deletions(-) diff --git a/docs/deployments/container/accessing-images.mdx b/docs/deployments/container/accessing-images.mdx index 3a793aed..2a2bd759 100644 --- a/docs/deployments/container/accessing-images.mdx +++ b/docs/deployments/container/accessing-images.mdx @@ -64,6 +64,12 @@ See [how to run the Core Speech CPU container here.](/deployments/container/cpu- The Transcription GPU images are required to use the most accurate models. +See [how to run the Transcription GPU container here.](/deployments/container/gpu-speech-to-text) + +:::info +To access additional language configurations for containers, please [speak to our Support Team](https://support.speechmatics.com). +::: + ### Standard model There is a single image available that supports all languages for the Standard model. There are language specific images available that support the Enhanced and Standard models. @@ -78,7 +84,7 @@ docker pull speechmaticspublic.azurecr.io/sm-gpu-inference-server-en:${smVariabl ### Enhanced model -Depending on which Enhanced model languages are required, you can pull specific images, or a single image covering all languages. +Depending on which Enhanced model languages are required, you can pull specific images.
Language Pack 1 @@ -112,21 +118,16 @@ Depending on which Enhanced model languages are required, you can pull specific
-
- All languages -

All Enhanced and Standard model languages

- - {`docker pull speechmaticspublic.azurecr.io/sm-gpu-inference-server-all-lang:${smVariables.latestContainerVersion} `} - -
+### Standard and Enhanced model -To load only some of an image's languages at startup, see [Loading only selected languages](/deployments/container/gpu-speech-to-text#loading-only-selected-languages). +A single image supports all languages for both the Standard and Enhanced models, so you do not need to pull individual language pack images. -See [how to run the Transcription GPU container here.](/deployments/container/gpu-speech-to-text) + + {`# pulling the Transcription GPU inference server supporting all languages for both Enhanced and Standard models with the ${smVariables.latestContainerVersion} tag: +docker pull speechmaticspublic.azurecr.io/sm-gpu-inference-server-all-lang:${smVariables.latestContainerVersion}`} + -:::info -To access additional language configurations for containers, please [speak to our Support Team](https://support.speechmatics.com). -::: +This image contains both models. To load only one of them, see [Running only one model](/deployments/container/gpu-speech-to-text#running-only-one-model). To load only some of the image's languages, see [Loading only selected languages](/deployments/container/gpu-speech-to-text#loading-only-selected-languages). ### Melia 1 model diff --git a/docs/deployments/container/gpu-speech-to-text.mdx b/docs/deployments/container/gpu-speech-to-text.mdx index de99cf78..2bb789a2 100644 --- a/docs/deployments/container/gpu-speech-to-text.mdx +++ b/docs/deployments/container/gpu-speech-to-text.mdx @@ -119,7 +119,7 @@ When running the all language standard model GPU inference server you must set t ### Loading only selected languages -Images that contain more than one language model, such as the [all language images](/deployments/container/accessing-images#enhanced-model), load every language at startup. To save GPU memory for throughput, pass the `SM_LANGUAGES` environment variable to the container and set it to a comma-separated list of language codes. Only those languages are loaded: +Images that contain more than one language model load every language at startup. To save GPU memory for throughput, pass the `SM_LANGUAGES` environment variable to the container and set it to a comma-separated list of language codes. Only those languages are loaded: ```bash -e SM_LANGUAGES=en,de,fr @@ -129,6 +129,8 @@ Codes must be languages the image contains. A code the image does not contain fa When `SM_LANGUAGES` is unset, all of the image's languages are loaded. `SM_LANGUAGES` is available from Container version 15.18.0. +`SM_LANGUAGES` and `SM_MODEL` can be combined to trim both axes of the startup load. For example, `SM_MODEL=standard` with `SM_LANGUAGES=en,de` loads the Standard model for English and German only. + ### Monitoring the server The inference server is based on [Nvidia's Triton architecture](https://developer.nvidia.com/nvidia-triton-inference-server) and as such From 6929876d6f6fe60671cd83d2954d3cc0d9dcb75d Mon Sep 17 00:00:00 2001 From: Venkatesh Chandran Date: Tue, 11 Aug 2026 10:30:28 +0100 Subject: [PATCH 3/4] Keep the batch benchmark version at 15.7.0 The benchmark figures were measured on 15.7.0, so the configuration table should continue to report that version. Co-Authored-By: Claude Opus 5 (1M context) --- .../container/performance-and-cost.mdx | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/docs/deployments/container/performance-and-cost.mdx b/docs/deployments/container/performance-and-cost.mdx index 7e9392f9..addd0525 100644 --- a/docs/deployments/container/performance-and-cost.mdx +++ b/docs/deployments/container/performance-and-cost.mdx @@ -24,15 +24,15 @@ This is a comparison of the performance and estimated running costs of transcrip The benchmark uses the following configuration: -| Benchmark details | | -|-------------------|-----------------------------------------------| -| Version | 15.18.0 (Standard, Enhanced), 1.3.0 (Melia-1) | -| Language | English only | -| CPU | D16ds_v5 | -| GPU Standard | Standard_NC16as_T4_v3 | -| GPU Enhanced | Standard_NC8as_T4_v3 | -| GPU Melia-1 | Standard_NC40ads_H100_v5 | -| Price Basis | Azure PAYG East US, Linux, Standard | +| Benchmark details | | +|-------------------|----------------------------------------------| +| Version | 15.7.0 (Standard, Enhanced), 1.3.0 (Melia-1) | +| Language | English only | +| CPU | D16ds_v5 | +| GPU Standard | Standard_NC16as_T4_v3 | +| GPU Enhanced | Standard_NC8as_T4_v3 | +| GPU Melia-1 | Standard_NC40ads_H100_v5 | +| Price Basis | Azure PAYG East US, Linux, Standard | :::note For GPU Models, transcribers and inference servers were all run on a single VM node. From 5120ff53fdf67f98ff6f9f7dade6a383a0d11e13 Mon Sep 17 00:00:00 2001 From: Venkatesh Chandran Date: Tue, 11 Aug 2026 12:01:30 +0100 Subject: [PATCH 4/4] Address review comments on the SM_LANGUAGES documentation - Lowercase "container version" per terminology guidance - Comment the SM_LANGUAGES example with the languages it loads - Reword the unrecognised-code sentence for grammar - Replace "trim both axes of the startup load" with plainer wording - Distinguish sm-gpu-inference-server-all-lang from the Standard-only sm-gpu-inference-server-standard-all image Co-Authored-By: Claude Opus 5 (1M context) --- docs/deployments/container/accessing-images.mdx | 2 +- docs/deployments/container/gpu-speech-to-text.mdx | 7 ++++--- 2 files changed, 5 insertions(+), 4 deletions(-) diff --git a/docs/deployments/container/accessing-images.mdx b/docs/deployments/container/accessing-images.mdx index 2a2bd759..fda15cae 100644 --- a/docs/deployments/container/accessing-images.mdx +++ b/docs/deployments/container/accessing-images.mdx @@ -120,7 +120,7 @@ Depending on which Enhanced model languages are required, you can pull specific ### Standard and Enhanced model -A single image supports all languages for both the Standard and Enhanced models, so you do not need to pull individual language pack images. +A single image supports all languages for both the Standard and Enhanced models, so you do not need to pull individual language pack images. The `sm-gpu-inference-server-standard-all` image above covers all languages for the Standard model only. {`# pulling the Transcription GPU inference server supporting all languages for both Enhanced and Standard models with the ${smVariables.latestContainerVersion} tag: diff --git a/docs/deployments/container/gpu-speech-to-text.mdx b/docs/deployments/container/gpu-speech-to-text.mdx index 2bb789a2..ccef0308 100644 --- a/docs/deployments/container/gpu-speech-to-text.mdx +++ b/docs/deployments/container/gpu-speech-to-text.mdx @@ -122,14 +122,15 @@ When running the all language standard model GPU inference server you must set t Images that contain more than one language model load every language at startup. To save GPU memory for throughput, pass the `SM_LANGUAGES` environment variable to the container and set it to a comma-separated list of language codes. Only those languages are loaded: ```bash +# load only English, German, and French -e SM_LANGUAGES=en,de,fr ``` -Codes must be languages the image contains. A code the image does not contain fails startup with an error. +Codes must be languages the image contains. Any code the image does not contain causes startup to fail with an error. -When `SM_LANGUAGES` is unset, all of the image's languages are loaded. `SM_LANGUAGES` is available from Container version 15.18.0. +When `SM_LANGUAGES` is unset, all of the image's languages are loaded. `SM_LANGUAGES` is available from container version 15.18.0. -`SM_LANGUAGES` and `SM_MODEL` can be combined to trim both axes of the startup load. For example, `SM_MODEL=standard` with `SM_LANGUAGES=en,de` loads the Standard model for English and German only. +`SM_LANGUAGES` and `SM_MODEL` can be used together. For example, `SM_MODEL=standard` with `SM_LANGUAGES=en,de` loads the Standard model for English and German only. ### Monitoring the server