Speech

The speech group configures the speech model (Qwen3-ASR, on the CPU, beside the model on the card). Every field is read live: PATCH /v1/config changes it with no reload, and threads and pinned_weights hold from the model’s next load. The generated list, with every default, is the configuration reference.

Flag	Env	Default	Meaning
--speech-model	IGNIS_SPEECH_MODEL	auto	auto is qwen3-asr-1.7b, the 1.7B model with a Q4_K_M decoder; a catalog id picks another entry; off turns the speech model off. It is on when its files are on disk (ignis-server model download <id>); nothing downloads by itself. A catalog entry that is not a speech model (a text or image model’s id) is refused, at start and by a PATCH.
--speech-compat-names	IGNIS_SPEECH_COMPAT_NAMES	whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-transcribe, gpt-live-transcribe	The model names OpenAI clients send that the speech model answers to (Aider hard-codes whisper-1). ;- or ,-separated. Not catalog aliases and not listed by GET /v1/models. A name equal to a served model id, or to an id or alias of the catalog in use (an operator’s download.catalog among it), is refused at start and by a PATCH.
--speech-threads	IGNIS_SPEECH_THREADS	auto	The CPU threads the model computes with; auto is the physical cores less one, counted within the CPUs the server may run on (a container’s cpuset or CPU quota). Each load logs the count and whether auto chose it (ignis.speech.threads).
--speech-max-clip-seconds	IGNIS_SPEECH_MAX_CLIP_SECONDS	1200	The longest clip one transcription takes (1..1200, the model’s limit). A longer one is a 400; nothing is split.
--speech-context-max-tokens	IGNIS_SPEECH_CONTEXT_MAX_TOKENS	512	The most tokens of prompt and keywords[] a transcription reads; past it the last are kept. 0 reads none.
--speech-pinned-weights	IGNIS_SPEECH_PINNED_WEIGHTS	off	Pin the weights in host memory instead of mapping the file: the model’s tensors are copied into the server’s own memory at load (and locked in RAM where the OS allows it, on Linux), so the files are not held open.
--speech-idle-unload	IGNIS_SPEECH_IDLE_UNLOAD	300	Seconds with no transcription, and no live session open, after which the model is unloaded; the next one loads it again. 0 never unloads.
--speech-max-queue	IGNIS_SPEECH_MAX_QUEUE	4	The transcriptions that may wait behind the one running; one more is 503 engine_full with Retry-After. 0 runs one at a time.
--speech-max-sessions	IGNIS_SPEECH_MAX_SESSIONS	4	The most live transcription sessions open at once; one more is refused at the upgrade, 503 engine_full with Retry-After.
--speech-partial-interval-ms	IGNIS_SPEECH_PARTIAL_INTERVAL_MS	2000	How often at most a live session’s transcript refreshes while audio arrives: each refresh sends the new stable text, and the provisional tail to a session that asked for it. Best effort: a refresh is skipped while the session’s audio is still being computed. Shorter is livelier text and more CPU per session.
--speech-partial-cost-factor	IGNIS_SPEECH_PARTIAL_COST_FACTOR	1	How long the next refresh of a live transcript waits at least, as a multiple of the last refresh’s compute time; 1 only waits for the last refresh to finish (the text).
--speech-rollback-tokens	IGNIS_SPEECH_ROLLBACK_TOKENS	10	The tokens of a live transcript held back as provisional, the rest being stable: the provisional tail a session that asks for it gets. Read by the speech executor when a session opens, so a change holds from the next session.
--speech-unfixed-partials	IGNIS_SPEECH_UNFIXED_PARTIALS	2	The first refreshes of a live session’s transcript that make no text stable, whatever speech.rollback_tokens says: the language and the opening words are decided on more than one interval of audio. 0 fixes text from the first refresh. Read when a session opens, so a change holds from the next session.
--speech-vad-threshold, --speech-vad-prefix-padding-ms, --speech-vad-silence-duration-ms	IGNIS_SPEECH_VAD_*	0.5, 300, 500	The defaults of a live session’s server_vad, OpenAI’s: the speech probability that counts as speech, the audio an utterance keeps from before it, and the silence after it at which the server commits it. A session reads them when it opens; one that sends its own values uses those.
--speech-vad-leading-silence-ms	IGNIS_SPEECH_VAD_LEADING_SILENCE_MS	500	The digital silence the speech model reads before each server_vad utterance, never shown to the client; it counts toward speech.max_clip_seconds (turn detection). 0 feeds the utterance as heard.
--speech-max-utterance-seconds	IGNIS_SPEECH_MAX_UTTERANCE_SECONDS	120	The length guard: once a server_vad utterance holds this many seconds, the server ends it at the next short pause, or utterance_cut_search_ms later where the audio is, and the speech goes on in a new item, so a room that never falls silent loses nothing to max_clip_seconds. A session’s ignis.max_utterance_seconds overrides it; 0 never cuts. At 240 s an utterance’s commit took four times as long as at 120 s, at no gain in word error rate.
--speech-utterance-cut-search-ms	IGNIS_SPEECH_UTTERANCE_CUT_SEARCH_MS	10000	How long past the limit the length guard looks for a pause before it cuts where the audio is. It looks forward only, so no text already sent is past the cut. In fluent speech the pause comes within 10 s 95 times in 100: between words the VAD’s speech probability seldom drops.
--speech-max-queued-audio-seconds	IGNIS_SPEECH_MAX_QUEUED_AUDIO_SECONDS	120	The audio a live session may hold that is not computed yet; an append past it is refused, ignis_audio_queue_full, and the session stays open (sending audio). Above the Playground’s 60 s pace for a file; provisional.
--speech-loan-grace-seconds, --speech-loan-wait-seconds	IGNIS_SPEECH_LOAN_*	30, 5	The loan of free KV pages to the speech model. Not served yet.

The fields marked “not served yet” are declared so that a configuration file written today keeps its meaning: the loan that reads them comes with its own ticket (ADR 0067).

The diarizer

The diarizer group configures the diarizer, Nemotron 3 Diarization, which gives a live session that asks for them its speakers. It runs on the CPU, on a thread and CPU threads of its own beside the speech model’s, so its steps never hold up the speech model’s work. Every field but diarizer.threads is read live, when a session’s speakers start; diarizer.threads is read by the next load. The diarizer is a catalog entry of its own, kind: diarizer (one GGUF, no sidecar): the built-in nemotron-3-diarization, fetched with ignis-server model download nemotron-3-diarization like any entry, or one of an operator’s catalog. It is loaded by the first session that asks for speakers, about 0.2 GB of RAM for its F16 weights, never VRAM, and unloaded with the speech model, after speech.idle_unload seconds with no session open.

What it costs, measured at the live geometry: on a 10-core i9-10900K a 2 s step takes about 0.63 s on its 3 threads alone (0.88 s on 2, 0.56 s on 4), and on a 297 s conversation with qwen3-asr-1.7b transcribing the same session at real time, 0.73 s on average, the speakers staying within 3.9 s of the audio, the commits no slower than with the diarizer sharing the speech model’s threads and the live text as frequent as without speakers; on 4 cores of a Linux container, the diarizer alone took 23 to 24 % of the four cores (3.2 to 4.1 times faster than real time), where the speech model’s refreshes alone already fall behind. A file, at the file geometry, costs 25 to 47 times less than its length. GET /v1/models says whether it can serve and whether it is loaded (the speech.diarizer member). The geometries and the speaker cache’s constants are Sortformer’s, Nemotron 3 Diarization’s own: a diarizer of another family would refuse them.

Flag	Env	Default	Meaning
--diarizer-model	IGNIS_DIARIZER_MODEL	auto	auto is the catalog’s diarizer entry; a catalog id picks another; off turns speakers off. A text, image or speech model’s id is refused, at start and by a PATCH /v1/config. A live change takes the next session’s load. A session open with speakers keeps the diarizer it opened with until it ends: meanwhile GET /v1/models reads the new one, and the old one’s file is not protected from model remove (on Windows its removal fails while the session holds it). Nothing downloads by itself.
--diarizer-threads	IGNIS_DIARIZER_THREADS	auto	The CPU threads the diarizer computes with, its own beside the speech model’s; auto is a third of the physical cores, at least 2 (3 on a 10-core machine), counted within the CPUs the server may run on. The load reads it, so a change holds from the next one, and logs the count (ignis.diarizer.load).
--diarizer-min-speaker-seconds	IGNIS_DIARIZER_MIN_SPEAKER_SECONDS	2	The seconds of speech a voice needs to be a speaker of its own; until then its speech counts for the speaker it sounds most like, so a short mistaken voice splits no utterance. 0 makes every voice a speaker at once. At 2 s a mistaken voice of a few seconds in test conversations went, and no participant of real meetings who spoke little was lost (at 5 s some were). The price: a new voice’s first seconds go to someone else (a colleague who at first only says “yes, exactly” is labelled as another speaker until they have spoken this long; the run of speech that makes them a speaker becomes theirs), and when the voice changes with no pause, the utterance before the change ends only once the new voice is a speaker.
--diarizer-live-speaker-cache-frames, -live-fifo-frames, -live-chunk-frames, -live-right-context-frames, -live-update-period-frames	IGNIS_DIARIZER_LIVE_*	264, 264, 25, 4, 222	The live geometry (ignis.diarization: "live"), in 80 ms frames: the speaker cache, the FIFO, the chunk a step scores, the look-ahead it waits for, and the frames an overflowing FIFO gives the cache. The chunk and the look-ahead are the latency, 2.32 s; this geometry measured a better DER than the card’s low-latency one at 2.6 times less CPU.
--diarizer-file-speaker-cache-frames, …	IGNIS_DIARIZER_FILE_*	264, 40, 340, 40, 300	The file geometry ("file"), the card’s offline one, for audio fed faster than real time: 30.4 s of latency, at 25 to 47 times less CPU than real time.
--diarizer-prediction-score-threshold, -latest-frames-score-boost, -strong-boost-rate, -weak-boost-rate, -min-positive-scores-rate, -speaker-cache-silence-frames-per-speaker	IGNIS_DIARIZER_*	auto	The speaker cache’s constants, by the checkpoint’s own names: how it scores and keeps the frames of each speaker. auto is the checkpoint’s value (0.25, 0.05, 0.75, 1.5, 0.5 and 1), which its weights carry; an override is bounded, and the values in force when the diarizer loads are logged (ignis.diarizer.load).