feat: add smallest ai asr and tts extensions#2203
Merged
plutoless merged 7 commits intoJul 18, 2026
Merged
Conversation
harshitajain165
force-pushed
the
feat/smallest-ai-integration
branch
from
July 7, 2026 07:11
63273f3 to
e2f2b42
Compare
harshitajain165
marked this pull request as ready for review
July 7, 2026 08:43
Contributor
Author
|
@halajohn @plutoless |
Contributor
Author
|
Hi @plutoless 👋 The earlier CI failure was in smallest_tts_python's test_headers_configuration (tests/test_params.py) — the client now sets a default X-Source: ten-framework header, so the header-count assertions were off by one. I've fixed the assertions and pushed the update. The workflow runs are currently sitting in "action_required" (the fork-PR approval gate), so they haven't started. Could you approve/re-run the CI workflows when you get a chance? And a review whenever you have time would be much appreciated. Thanks a lot! |
harshitajain165
force-pushed
the
feat/smallest-ai-integration
branch
2 times, most recently
from
July 14, 2026 09:32
d17c5d9 to
3f8eb1a
Compare
diyuyi-agora
previously approved these changes
Jul 16, 2026
Contributor
|
format check failed， please fix |
Add two new vendor extensions for Smallest AI: - smallest_asr_python: real-time speech-to-text via the Pulse live WebSocket API (binary PCM16 in, interim/final transcripts out, finalize control message, reconnect with exponential backoff). - smallest_tts_python: text-to-speech via the Lightning SSE streaming endpoint (base64 PCM16 chunks decoded on the fly, ~100 ms to first audio, output_format pinned to pcm). Both include mocked extension-local test suites, test configs, and README docs. Configured via params.api_key or SMALLEST_API_KEY.
…sults - guard _send_finalize against a down websocket and complete the finalize handshake locally so asr_finalize_end is always emitted and the turn cannot stall - reset the reconnect retry budget on received transcriptions instead of right after the websocket handshake, so an accept-then-close failure hits the max-attempts fatal error instead of looping forever - reformat _handle_error_message to be black-clean
- tts: stop substring-matching 401/403 in exception text; real auth failures are already classified by the HTTP status-code branch, so transport-level exceptions are always plain errors now - asr: latch the utterance start when its first transcript arrives and reuse it for word-less results, so interim start_ms stays monotonic instead of pointing at the end of the audio stream
The client now sets a default X-Source: ten-framework header, so the header-count assertions in test_headers_configuration were off by one.
…ancel When the vendor WebSocket dropped mid-session, _process_messages (running inside _message_task) awaited _handle_reconnect, whose start_connection -> stop_connection cancels _message_task itself, interrupting the reconnect before a new message task was created. Route every reconnect trigger through a _schedule_reconnect helper that runs the attempt on an independent, tracked task, so the message loop can exit cleanly while the reconnect proceeds. In-flight reconnect tasks are retained against GC and cancelled on deinit. Add a regression test that drops the socket mid-session and asserts the reconnect delivers a transcript.
After a failed reconnect self.ws can be None; the assert would raise (or be stripped under -O) instead of returning False gracefully. Fold the None check into the is_connected() guard, mirroring _send_finalize and pipecat's state-is-OPEN send guard.
harshitajain165
force-pushed
the
feat/smallest-ai-integration
branch
from
July 17, 2026 12:27
8d0f1f4 to
38f6a55
Compare
Contributor
Author
Could you re-run the CI workflows now? |
plutoless
approved these changes
Jul 18, 2026
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
3 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.
What
Adds two new vendor extensions for Smallest AI:
smallest_asr_python— real-time speech-to-text using thePulse
live WebSocket API (
wss://api.smallest.ai/waves/v1/stt/live). 38 languages,64 ms time-to-first-transcript.
smallest_tts_python— text-to-speech using theLightning
SSE streaming endpoint (
/waves/v1/tts/live). ~100 ms to first audio chunk,12 languages, voice cloning;
lightning_v3.1andlightning_v3.1_promodels.Implementation notes
ASR (extends
AsyncASRBaseExtension):final=false, finals onPulse's
is_final.asr_finalizemaps to Pulse's{"type": "finalize"}control message (sessionstays open).
word_timestamps=trueby default so final results carry accuratestart_ms/duration_ms.TTS (extends
AsyncTTS2HttpExtension):data: {"audio": "<base64>"}/{"done": true}) decodedincrementally; a partial line is never split across chunk boundaries.
output_formatpinned topcm: Lightning's raw PCM is already signed 16-bitLE mono, matching the
pcm_framecontract with no conversion, and skipping thecontainer header keeps time-to-first-audio low.
invalid_api_keyclassified asINVALID_KEY_ERROR(fatal);everything else non-fatal.
Both authenticate via
params.api_key/SMALLEST_API_KEYand sendX-Source: ten-frameworkfor API-side attribution. All otherparamskeys passthrough verbatim (query string for ASR, request body for TTS).
Testing
ASR — result shape, finalize, dump, metrics, reconnect, vendor error, invalid
params; TTS — basic/dump/flush, error classification, metrics, params/URL
resolution, robustness, state machine.
python -m black --check --line-length 80clean; syntax verified.task asr-guarder-test EXTENSION=smallest_asr_python— 10/10 passed against the live APItask tts-guarder-test EXTENSION=smallest_tts_python— 15/15 passed (+1 optional skip) against the live APIvoice_assistant_smallestgraph (Pulse STT + GPT-4o + Lightning TTS over Agora RTC), including barge-in/interruptionSupersedes #2200 (auto-closed by a force-push during a history cleanup).