Model configurations¶
The raw results behind the notes below are in the vendor notes; vxasr is described on its own page.
A configuration is a provider, a model, and the post-processing chain applied to it. Post-processing is part of a configuration’s identity rather than a request flag, so a model raw and the same model with Groq formatting are two separately selectable configurations that can be compared on equal terms:
Configuration id |
Needs |
|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
nothing |
Most Qwen model ids are dated snapshots, not floating ones — the vendor repoints undated ids without notice, which would make a vote name a moving target. The exceptions are qwen-omni/qwen3.8-omni-flash-realtime and the two qwen-audio/* ids, which the vendor lists with no snapshot id yet. The Qwen Omni models need no +groq variant: their output is already well-formatted (Thai words in Thai, product names in Latin), which is what Groq formatting was tidying up for the plain ASR models — running it on top of Omni’s output measurably added nothing.
BytePlus exposes two modes at two endpoints, and only bigmodel_nostream accepts a language — the bi-directional bigmodel mode covers Mandarin, English and a few Chinese dialects and nothing else, so on any other language it returns confident nonsense. Only bigmodel_nostream is offered as a configuration for that reason; the provider still serves bigmodel for deployments in those languages. Set BYTEPLUS_LANGUAGE (e.g. th-TH) or BytePlus falls back to that same Chinese/English default set. Note that bigmodel_nostream returns results only after 15 s of audio or the final packet, so it emits few or no partials — it is accuracy-tuned rather than low-latency.
OpenRouter is a batch HTTP endpoint fanning out to many backing vendors under one key, not a live socket like every other provider here — the whole clip is sent as one file once recording stops, so it produces no partial transcript. Four models are declared, each tried live on the same fixture testdata/OBSERVATIONS.md uses. microsoft/mai-transcribe-2 is the default: when the three were finally timed against each other rather than compared on transcript alone, it was the fastest and the cheapest, and its transcript matched the best of them. meta/muse-voice-transcribe-1.0 held the default before that on transcript quality, and turned out to be four times slower than the others; microsoft/mai-transcribe-1.5 was the original pick from a comparison of 19 models. google/gemini-3.5-transcribe is the Gemini model described below, called as a batch request.
The meta provider speaks to that same muse-voice-transcribe-1.0 model directly over Meta’s own realtime WebSocket endpoint instead of through OpenRouter’s batch wrapper, so it is the only configuration of this model that actually streams partial transcripts. Live-tried against the same fixture (testdata/OBSERVATIONS.md): the transcript came back slightly less accurate than the OpenRouter route for this same model on that one clip, and usage is billed in whole seconds rounded up rather than OpenRouter’s fractional cost — a genuine tradeoff, not a strict upgrade, which is why both configurations are kept.
Paxa Labs is a batch endpoint too, but takes JSON with the audio base64-encoded rather than a multipart upload. It is the only model compared in testdata/OBSERVATIONS.md that needed no instruction to behave: no filler tags, no spaces inserted between Thai words, no English translated into Thai, and byte-identical output across repeated runs of every clip. It costs roughly four times mai-transcribe-2 for near-identical speed, so it is offered alongside rather than as a default. PAXA_CONVENTION=written renders numbers as digits and uses the Thai repetition mark; the vendor’s default spells them as spoken.
paxa/paxa-stt-lite-realtime-v1-preview is Paxa’s realtime WebSocket. The vendor detects the end of speech, so the final transcript arrives about 55 ms after the recording ends. It costs 1.5 times the batch rate, and the whole connection is charged, including silence.
Gemini 3.5 Transcribe is offered two ways. Through OpenRouter it is a batch call: accurate and cheap, but 4–6 s per request whatever the clip length. The gemini provider uses the Live model instead, in push-to-talk mode with smart transcription (fillers and false starts removed): partial transcripts about every 0.5 s, and the final 0.3–0.6 s after the audio ends. A Live session can stream for 10 minutes at most. In tests on clips joined from separate recordings, push-to-talk with smart transcription sometimes dropped a later sentence; see testdata/OBSERVATIONS.md.
ai-gateway/microsoft/mai-transcribe-2-streaming is the streaming version of MAI-Transcribe-2, offered only through the Vercel AI Gateway, and reached with the AI SDK’s experimental_streamTranscribe. Partial transcripts come about every 0.3 s, and the final 0.33–0.49 s after the audio ends. On clips with pauses between sentences it kept every sentence. It costs $0.54 per hour of audio.
Ids contain +, which decodes to a space in a query string, so clients must URL-encode them — URLSearchParams does this automatically. Several ids also contain / (for example openrouter/microsoft/mai-transcribe-1.5), which needs no special handling in a query string value.
By default a configuration is selectable once the environment carries every credential it needs — which means mock/mock, needing none, is always selectable. Set ASR_CONFIGURATIONS to pin the list down explicitly. Requests naming an unknown or disabled configuration are closed with code 1008, and one named without its credentials with 1011 (a server misconfiguration), before any message is created.
Discovering them¶
GET /asr/configurations returns the selectable set, so a client does not have to hardcode the table above:
{
"primaryConfigurationId": "qwen-omni/qwen3.5-omni-flash-realtime-2026-03-15",
"configurations": [
{
"id": "qwen-omni/qwen3.5-omni-flash-realtime-2026-03-15",
"label": "Qwen3.5-Omni-Flash Realtime (2026-03-15, raw)",
"providerId": "qwen-omni",
"model": "qwen3.5-omni-flash-realtime-2026-03-15",
"postProcessing": [],
"configured": true
}
]
}
Every listed id is one /ws accepts in ?configuration=, which is what makes this list usable as a fan-out plan: an eval opens one socket per entry. primaryConfigurationId names the configuration the primary path uses; it is listed alongside the rest, since it competes on the same terms.
configured is almost always true — a configuration normally earns its place here by being credentialled. It is false only where an ASR_CONFIGURATIONS entry (or the default, which is always listed) was named without its credentials; such an entry is shown rather than hidden so the gap is visible, but connecting to it would close immediately. The response never carries credentials — not their values, and not the names of the missing variables either; the /ws close reason is where an operator reads which one to set.