feat: SDK 7.7.0 — Flux TTS controls, Flux STT fix, listen redact/numerals - #92
Open
GregHolmes wants to merge 16 commits into
Open
feat: SDK 7.7.0 — Flux TTS controls, Flux STT fix, listen redact/numerals#92GregHolmes wants to merge 16 commits into
GregHolmes wants to merge 16 commits into
Conversation
Bump deepgram-sdk 7.5.0 -> 7.7.0 across the workspace and expose the new Flux (Speak v2) streaming controls on `dg speak`: - `--speed` (0.85-1.15, 0.05 steps) and `--expressivity` (-2..2) forward to `speak.v2.connect()` via `speak_text_stream()`; only sent when set. - Both are validated up front and rejected for Aura (v1) models, so misuse fails with a clear message instead of a mid-stream server error. - Update --help, examples, agent_help, the skill-generator snippet, README, and add tests (forwarding, Aura rejection, invalid speed/expressivity).
Expose --redact (numbers, aggressive_numbers, or v1 values like pci/ssn) and --numerals on `dg listen`, applied to both prerecorded and live paths. Also fixes Flux STT (listen v2) streaming, which was broken since it was added: - `_ws_url` sent v1-only params (language, smart_format, punctuate, channels, diarize, interim_results) to the v2 endpoint, which rejected them with HTTP 400. Build the param set per API version instead. - The message parser only understood v1 `Results`; Flux emits turn-based `TurnInfo` events. Add `_handle_v2_turn` with per-turn state, plus `_flush_v2` to emit the final turn when a finite stream closes before an `EndOfTurn` (common for files/stdin). Verified against staging: `--model flux-general-en --numerals` yields digit transcripts, `--redact numbers` redacts them, and nova-3 (v1) is unaffected. Adds tests for _ws_url params and v2 turn finalization.
SDK 7.6.0 added Aura-2 Spanish voices (e.g. aura-2-selena-es). These need no CLI code change — the model is passed through as a string — so just add an example and a pointer to `dg models` for the current list.
The Flux STT (listen v2) HTTP 400 fix hinges on _ws_url NOT sending v1-only params (language, smart_format, punctuate, channels, diarize, interim_results) to the v2 endpoint, but nothing asserted that directly — a revert would silently rebreak Flux STT. Add explicit v2-omits / v1-includes assertions.
Add tests/e2e/ that drive command handle() against the real Deepgram API in-process, covering the release's new functionality end to end: - speak Flux --speed/--expressivity -> valid WAV - speak aura-2-selena-es (multilingual) -> MP3 over REST - listen Flux STT (v2) --numerals -> digits in transcript - listen Flux STT (v2) --numerals --redact numbers -> redacted - listen nova-3 (v1) baseline -> guards the v2-fix regression Gated on DEEPGRAM_API_KEY (skipped otherwise), so they never run in the CI matrix (no secret) but run locally/manually with a key exported. Set DEEPGRAM_BASE_URL to target staging. Verified: 5 skip without a key, 5 pass against production.
…(v2) Address review findings on PR #92: - listen: emit a stderr note when --diarize is set against a Flux STT (v2) model, instead of silently dropping it. v1 paths unaffected. - disambiguate user-facing copy now that both Flux surfaces exist: "Flux TTS" in speak, "Flux STT" in listen (help text, the Aura rejection error, and README comments). - bound the SDK pin to >=7.7.0,<8 across all packages so a future breaking major isn't pulled in automatically. Gate: make check clean; full suite + live e2e (production) green.
Two upfront guards in `listen` so Flux (v2) misuse fails with a clear
message instead of an opaque server HTTP 400:
- Reject `--redact` values outside {numbers, aggressive_numbers} on Flux
models, mirroring speak's Flux-flag validation.
- Reject Flux model + pre-recorded file/URL (v2 is streaming-only) before
it routes to /v1/listen and surfaces a wrapped header dump.
Also reformat test_speak_command.py to satisfy ruff format.
The v1 API accepts multiple redact categories (redact=pci&redact=numbers) but the flag was single-valued. Mark --redact multiple=True and normalise to a tuple. - REST path passes a list so Fern's query encoder expands it into repeated params (a tuple is left unexpanded). - WebSocket path uses urlencode(doseq=True) to expand the sequence. - Flux (v2) validation now checks every supplied value against the numbers/aggressive_numbers vocabulary.
The S2 guard now blocks flux model + file before dispatch, so the two TestFluxModelAutoVersion tests that asserted flux+file reaching _prerecorded with api_version=2 crashed (mock never called). Split them to match the new behaviour: flux+file now asserts the streaming-only error result, and the flux→v2 assertion moves to the mic streaming path (which is where a Flux model is actually valid).
Rewrite dict() calls as literals (C408) and rename ambiguous `l` loop variable to `line` (E741). No behaviour change.
…rrors
Review of the Flux STT work surfaced three issues:
- Captions crashed / mis-timed on Flux STT. TurnInfo words carry only
{word, confidence} (no per-word start/end), so the live cue span fell
back to 00:00:00 and the end-of-stream batch save raised
KeyError: 'start'. Key caption timing off the turn's guaranteed
audio_window_start/end and backfill per-word timings from it so both the
live cues and captions_from_words() produce valid output.
- Flux STT fatal errors (control frame type "Error") were silently
dropped, leaving the user with an empty transcript. Surface code +
description to stderr.
- speak: note that --expressivity is beta and fixed for the connection.
Adds unit coverage for the audio-window caption timing (incl. the batch
save that used to KeyError) and the error-frame surface.
Close the two coverage gaps left from the Flux STT review: - Unit: assert _ws_mic catches a KeyboardInterrupt at the gather await, runs _flush_v2, and returns the in-flight turn's transcript (rather than losing the final partial turn on Ctrl-C). - Live e2e: Flux STT (v2) --srt captions must carry well-formed, non-zero timestamps (keyed off TurnInfo.audio_window_*), guarding the caption fix and the batch-save path that previously raised KeyError: 'start'.
…n edges) Fill the unit-coverage gaps the review surfaced in the new Flux/redact logic: - _prerecorded builds redact as a list + numerals "true" (Fern repeats the query param) — previously only the routing-to-_prerecorded was tested. - _timed_v2_words: synth-on-empty, window-spread, and real-timing passthrough. - v2 Update interim print, and EndOfTurn with an empty transcript (no-op). All new logic paths now unit-covered; the only remaining misses in command.py are pre-existing branches and the real mic/stdin/websocket I/O loops, which the live e2e suite exercises.
dg-coreylweathers
approved these changes
Aug 14, 2026
dg-coreylweathers
left a comment
dg-coreylweathers
left a comment
There was a problem hiding this comment.
Approved — verified via two independent review passes (details in review notes). All prior findings resolved at 074090d; CI green 38/38; live-verified against production. Three small non-blocking should-fixes noted for follow-up: (1) bare '-m flux' on dg speak exits 0 on failure, (2) invalid --redact on v1 prerecorded is silent + exit 0, (3) Ctrl-C on 'dg listen -' drops the transcript/--save-to output that the mic path preserves. Please confirm publish ordering: core 0.2.15 must land before speak/listen/root.
… on Ctrl-C Three error-contract fixes surfaced in review: - speak: the REST (Speak v1) handler returned BaseResult(status="error") instead of raising, which the framework prints but exits 0. Now that a bare/typo'd `flux` model routes to REST (not the raising v2 path), its failure regressed to exit 0. Raise ClickException so it exits non-zero. - listen: the prerecorded REST handler had the same return-not-raise sink; a --redact value the v1 endpoint refuses printed nothing and exited 0. Raise instead so the rejection is visible and non-zero. - listen: `dg listen -` (stdin) dropped the partial transcript and --save-to on Ctrl-C because _ws_stdin re-raised the interrupt. Mirror the mic path: catch KeyboardInterrupt/CancelledError, drain, and fall through to the flush + return so the transcript survives (fatal errors still propagate). Regression tests: REST API error exits non-zero (speak flux/aura, listen v1); stdin Ctrl-C preserves transcript and writes --save-to. Full suite 1037 passed.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Brings the CLI up to date with Deepgram Python SDK 7.6.0 + 7.7.0 in one release. The API-facing changes are additive and existing v1 behavior remains covered.
SDK and package constraints
deepgram-sdkis bounded at>=7.7.0,<8across the workspace.>=0.2.15, Speak>=0.0.4, and Listen>=0.0.14.uv.lockis intentionally gitignored; localuv lock --checkresolves 143 packages.Speak — Flux TTS streaming controls (7.7.0)
dg speak --speed(0.85–1.15, 0.05 steps) and beta--expressivity(-2..2), forwarded tospeak.v2.connect()only when set.flux-*models.flux-boundary; bare names and typos stay on the v1 pass-through path.send_interrupt()/send_configure()barge-in remain intentionally unexposed because they target live agent pipelines, not one-shotdg speak.Listen — redact / numerals + Flux STT fix (7.6.0 + 7.7.0)
dg listen --redactand--numerals, applied to prerecorded and live paths.numbers/aggressive_numbersvocabulary up front; v1 continues to accept categories such aspciandssn.language,smart_format,punctuate,channels,diarize,interim_results) that caused HTTP 400.TurnInfoevents are assembled per turn and finite streams flush the final in-flight turn.--diarizewarns and is cleared consistently from forwarding, UI, and result metadata.Docs
dg models.Live e2e safety
RUN_LIVE_E2E=1plus either an explicitDEEPGRAM_BASE_URLorRUN_LIVE_E2E_PRODUCTION=1.Verification
make check: Ruff format/check + mypy clean across 115 source files.uv lock --check: 143 packages resolved.Production behavior was verified during development for Flux TTS controls, Aura Spanish, Flux STT numerals/redaction, nova-3 v1, and Flux SRT timestamps. Version and changelog generation remain owned by release-please.