Overview — ddx-cuda-live-stt

ddx-cuda-live-stt

Streaming ASR · NVIDIA Parakeet-TDT-0.6B-v3 · :4600 · Deepgram-shaped wire format (v1.0.0 frozen).

Recommended SDK drop-in: point the official @deepgram/sdk base URL at the bare host. The SDK appends /v1/listen itself. Send a Deepgram credential (Authorization: Token <key> header, or the token, <key> subprotocol in the browser) and the connection routes to the Deepgram-compatible session. With no credential it falls through to the native handler. /v1/listen/dg and /deepgram/v1/listen are explicit aliases — not what you configure in the SDK.

Auth forms

Form	Wire	Used by
Header	`Authorization: Token <key>` (capital T, single space)	server SDK (Node, Python)
Subprotocol	`Sec-WebSocket-Protocol: token, <key>`	browser SDK (cannot set headers)

Form

Wire

Used by

Header

Authorization: Token <key> (capital T, single space)

server SDK (Node, Python)

Subprotocol

Sec-WebSocket-Protocol: token, <key>

browser SDK (cannot set headers)

Audio format

16 kHz mono linear16 (int16 LE). Chunk 3200 bytes (=100 ms) every ~50 ms so live VAD/endpointing runs. Other rates/channels are resampled/downmixed server-side.

Frame sequence

On a Deepgram session: Metadata (open) → Results (interim + final) → UtteranceEnd (on VAD speech-end) → Metadata (final, on CloseStream). SpeechStarted + UtteranceEnd require vad_events=true.

Query params (live registry)

Full table with behaviour badges + notes: /v1/help/params. Rendered live from the dispatch registry — never drifts.

Param	Behaviour	Note
`encoding`	SUPPORTED	Parakeet expects PCM-16 (linear16); transcode mulaw/alaw/opus.
`sample_rate`	SUPPORTED	Parakeet trained at 16 kHz; resample any declared rate.
`channels`	PARTIAL	Accept up to 2; mix to mono internally.
`endpointing`	SUPPORTED	Silero/WebRTC VAD; emit speech_final after silence>endpointing ms.
`model`	PARTIAL	Accept any value, route to Parakeet, echo in metadata.model_info.name.
`version`	NOOP	Legacy Deepgram field; accept silently, ignore.
`tier`	NOOP	Legacy Deepgram field; accept silently, ignore.
`language`	SUPPORTED	Parakeet-TDT-0.6B-v3 supports 25 languages; pass as model hint.
`punctuate`	PARTIAL	Parakeet basic punctuation; post-processor for accuracy.
`profanity_filter`	NOOP	Accepted and echoed; the transcript is returned unfiltered. No profanity filtering is applied.
`redact`	NOOP	Accepted and echoed; the transcript is returned unredacted. No redaction is applied — do not rely on this param for PII handling.
`diarize`	PARTIAL	Baseline speaker:0. When DDX_STT_DIARIZE_ENABLED=1, real Sortformer speaker labels on words[] (en supported; de/fr/es EXPERIMENTAL).
`multichannel`	ERROR	Reject if true with channels>1: close 1008 'multichannel not supported'.
`numerals`	SUPPORTED	Parakeet TDT head includes ITN; enable via model config.
`smart_format`	PARTIAL	Baseline: punctuate + numerals. When DDX_STT_SMART_FORMAT_ENABLED=1, engages WFST ITN on the FINAL transcript (en production; de/fr/es EXPERIMENTAL).
`replace`	PARTIAL	Repeatable 'search:replace'; simple string-replace post-processor.
`keywords`	SUPPORTED	POST-TRANSCRIPTION ONLY: repeatable 'kw[:intensifier]' is parsed, stored on the session, and used to drive phonetic + fuzzy correction of the FINAL transcript against the locale medical lexicon (Cologne Phonetics for de, Double Metaphone for en/fr/es). Interim results are NEVER rewritten, so live captions do not flicker. This is NOT acoustic/decoder-level biasing: the decode is byte-identical with or without it, because the service stays pinned to decoding strategy 'greedy' to dodge an NVRTC CUDA-graph capture bug (services/nemo_decoding.py). A 2026-08-03 spike found that pin's blocker gone but did NOT demonstrate a boosting tree loading (fusion_models stayed None), so true acoustic boosting remains unproven work. The intensifier suffix is accepted for Deepgram wire compatibility but does not weight the correction. Gated by DDX_STT_TERM_CORRECTION_ENABLED.
`keyterm`	NOOP	ACCEPTED BUT INERT: Deepgram's Nova-3 keyterm surface (max 500 tokens) is parsed and echoed, but has no effect on the transcript — same decoder-side gap as 'keywords'. Not implemented, no acoustic or post-processing effect.
`interim_results`	SUPPORTED	Stream partial Results frames; if false buffer to final only.
`utterances`	PARTIAL	Per-utterance segmentation via VAD + utterance_end_ms silence.
`utt_split`	PARTIAL	Silence threshold seconds; map to VAD min_silence_ms.
`utterance_end_ms`	SUPPORTED	Silence (ms) after last word before UtteranceEnd fires; drives SDK end-of-speech timing. None falls back to VAD default.
`no_delay`	NOOP	DG smart_format no-delay flag; our streaming path emits without an endpointing buffer delay already, so honour-but-noop (echoed).
`turn_detection`	PARTIAL	Dudoxx ext (shard 06). Silero endpointing is default. When DDX_STT_TURN_DETECTOR_ENABLED=1, a semantic EOU model REFINES each Silero speech_end (true turn-end vs mid-utterance pause); off = Silero only.
`turn_decision`	PARTIAL	Dudoxx ext. When DDX_STT_TURN_DECISION_ENABLED=1, every is_final Results frame carries ddx_turn {decision: continue\|complete\|incomplete, status, source: llm\|livekit, ...} — a dudoxx-LLM verdict over the turn text since the last complete, LiveKit EOU fallback (2-state). Advisory; never re-times a finalize.
`vad_events`	SUPPORTED	Silero/WebRTC VAD; emit SpeechStarted + UtteranceEnd frames.
`tag`	NOOP	Repeatable opaque analytics tag; accept silently, log only.
`extra`	NOOP	Opaque echo-back string; surface in opening Metadata transaction_key.
`mip_opt_out`	NOOP	Model-improvement opt-out; honour silently (no audio retention anyway).
`callback`	NOOP	Async callback URL; not applicable to streaming WS.
`callback_method`	NOOP	POST/GET/PUT/DELETE for async callback; ignored in streaming.
`words`	SUPPORTED	Parakeet provides word-level timestamps natively.
`paragraphs`	NOOP	Accept; omit paragraph structure from response.
`summarize`	NOOP	Invariant 8: NEVER call server-side LLM; return null/empty.
`detect_language`	NOOP	Invariant 8: NEVER call server-side LLM; return null languages.
`detect_topics`	NOOP	Invariant 8: NEVER call server-side LLM; return empty topics.
`detect_entities`	NOOP	Invariant 8: NEVER call server-side LLM; return entities:[].
`dictation`	NOOP	Spoken commands as regular words; no command interpreter.
`search`	NOOP	Repeatable; omit search_results from response.
`filler_words`	PARTIAL	Parakeet drops fillers by default; toggling needs model-config change.
`alternatives`	NOOP	Always return 1 alternative; SDK handles missing N-best gracefully.
`authorization`	NOOP	SDK per-request key override; the WS handler reads Authorization header.

Param

Behaviour

Note

encoding

SUPPORTED

Parakeet expects PCM-16 (linear16); transcode mulaw/alaw/opus.

sample_rate

SUPPORTED

Parakeet trained at 16 kHz; resample any declared rate.

channels

PARTIAL

Accept up to 2; mix to mono internally.

endpointing

SUPPORTED

Silero/WebRTC VAD; emit speech_final after silence>endpointing ms.

model

PARTIAL

Accept any value, route to Parakeet, echo in metadata.model_info.name.

version

NOOP

Legacy Deepgram field; accept silently, ignore.

tier

NOOP

Legacy Deepgram field; accept silently, ignore.

language

SUPPORTED

Parakeet-TDT-0.6B-v3 supports 25 languages; pass as model hint.

punctuate

PARTIAL

Parakeet basic punctuation; post-processor for accuracy.

profanity_filter

NOOP

Accepted and echoed; the transcript is returned unfiltered. No profanity filtering is applied.

redact

NOOP

Accepted and echoed; the transcript is returned unredacted. No redaction is applied — do not rely on this param for PII handling.

diarize

PARTIAL

Baseline speaker:0. When DDX_STT_DIARIZE_ENABLED=1, real Sortformer speaker labels on words[] (en supported; de/fr/es EXPERIMENTAL).

multichannel

ERROR

Reject if true with channels>1: close 1008 'multichannel not supported'.

numerals

SUPPORTED

Parakeet TDT head includes ITN; enable via model config.

smart_format

PARTIAL

Baseline: punctuate + numerals. When DDX_STT_SMART_FORMAT_ENABLED=1, engages WFST ITN on the FINAL transcript (en production; de/fr/es EXPERIMENTAL).

replace

PARTIAL

Repeatable 'search:replace'; simple string-replace post-processor.

keywords

SUPPORTED

POST-TRANSCRIPTION ONLY: repeatable 'kw[:intensifier]' is parsed, stored on the session, and used to drive phonetic + fuzzy correction of the FINAL transcript against the locale medical lexicon (Cologne Phonetics for de, Double Metaphone for en/fr/es). Interim results are NEVER rewritten, so live captions do not flicker. This is NOT acoustic/decoder-level biasing: the decode is byte-identical with or without it, because the service stays pinned to decoding strategy 'greedy' to dodge an NVRTC CUDA-graph capture bug (services/nemo_decoding.py). A 2026-08-03 spike found that pin's blocker gone but did NOT demonstrate a boosting tree loading (fusion_models stayed None), so true acoustic boosting remains unproven work. The intensifier suffix is accepted for Deepgram wire compatibility but does not weight the correction. Gated by DDX_STT_TERM_CORRECTION_ENABLED.

keyterm

NOOP

ACCEPTED BUT INERT: Deepgram's Nova-3 keyterm surface (max 500 tokens) is parsed and echoed, but has no effect on the transcript — same decoder-side gap as 'keywords'. Not implemented, no acoustic or post-processing effect.

interim_results

SUPPORTED

Stream partial Results frames; if false buffer to final only.

utterances

PARTIAL

Per-utterance segmentation via VAD + utterance_end_ms silence.

utt_split

PARTIAL

Silence threshold seconds; map to VAD min_silence_ms.

utterance_end_ms

SUPPORTED

Silence (ms) after last word before UtteranceEnd fires; drives SDK end-of-speech timing. None falls back to VAD default.

no_delay

NOOP

DG smart_format no-delay flag; our streaming path emits without an endpointing buffer delay already, so honour-but-noop (echoed).

turn_detection

PARTIAL

Dudoxx ext (shard 06). Silero endpointing is default. When DDX_STT_TURN_DETECTOR_ENABLED=1, a semantic EOU model REFINES each Silero speech_end (true turn-end vs mid-utterance pause); off = Silero only.

turn_decision

PARTIAL

Dudoxx ext. When DDX_STT_TURN_DECISION_ENABLED=1, every is_final Results frame carries ddx_turn {decision: continue|complete|incomplete, status, source: llm|livekit, ...} — a dudoxx-LLM verdict over the turn text since the last complete, LiveKit EOU fallback (2-state). Advisory; never re-times a finalize.

vad_events

SUPPORTED

Silero/WebRTC VAD; emit SpeechStarted + UtteranceEnd frames.

tag

NOOP

Repeatable opaque analytics tag; accept silently, log only.

extra

NOOP

Opaque echo-back string; surface in opening Metadata transaction_key.

mip_opt_out

NOOP

Model-improvement opt-out; honour silently (no audio retention anyway).

callback

NOOP

Async callback URL; not applicable to streaming WS.

callback_method

NOOP

POST/GET/PUT/DELETE for async callback; ignored in streaming.

words

SUPPORTED

Parakeet provides word-level timestamps natively.

paragraphs

NOOP

Accept; omit paragraph structure from response.

summarize

NOOP

Invariant 8: NEVER call server-side LLM; return null/empty.

detect_language

NOOP

Invariant 8: NEVER call server-side LLM; return null languages.

detect_topics

NOOP

Invariant 8: NEVER call server-side LLM; return empty topics.

detect_entities

NOOP

Invariant 8: NEVER call server-side LLM; return entities:[].

dictation

NOOP

Spoken commands as regular words; no command interpreter.

search

NOOP

Repeatable; omit search_results from response.

filler_words

PARTIAL

Parakeet drops fillers by default; toggling needs model-config change.

alternatives

NOOP

Always return 1 alternative; SDK handles missing N-best gracefully.

authorization

NOOP

SDK per-request key override; the WS handler reads Authorization header.