Skip to main content
Speech

Speech to Text

Transcribe audio to text across 47 STT models — Indic-first saaras, Cartesia Ink, and Deepgram Nova.

Overview

Transcribe an audio file into text with any of 47 speech-to-text models. The default, saaras:v4, covers 22 Indian languages plus English with automatic language detection — but the model field takes any file-transcription STT id, so you can pick per request.

Endpoint: POST /v1/audio/transcriptions

Picking a model

If you needUseWhy
Indian languages, or code-mixed Hinglishsaaras:v4Purpose-built for Indic phonetics; 23 languages, all five output modes
The widest language coverageink-whisper100 languages, and cheaper per hour than the Indic models
English call-centre audiodeepgram-nova-3Strong on accented and noisy telephony English
Clinical or pharmaceutical audiodeepgram-nova-3-medical, deepgram-nova-3-pharma or scribe_v2_medicalTuned for medical terms and drug names; deepgram-nova-3-pharma is English only; scribe_v2_medical hears 100 languages
Highest accuracy on recorded audio, any of 100 languagesscribe_v2ElevenLabs Scribe v2: automatic language detection, 13 Indian languages — see ElevenLabs Scribe models
Long recordings, up to 2 hours per file (preview)mai-transcribe-260 languages with automatic detection — see MAI-Transcribe-2
Turn detection built into the modelink-2, assemblyai-universal-3.6-pro or deepgram-flux-general-enVoice sessions only — see the note below

One model is on the free tier: saaras:v4. The retired nova-3 and whisper-large-v3-turbo IDs still work and stay free; they are answered by deepgram-nova-3 and whisper — see Retired and deprecated models.

The retired nova-3 ID and deepgram-nova-3 are the same model at different prices: nova-3 stays on the free plan at its old rate. Full per-model pricing: Models.

How transcription works

  1. 1

    Your app

    Upload an audio file (WAV/MP3) to

  2. 2

    CallMissed gateway

    Validate the key, resolve , detect language (or use ), apply

  3. 3

    Your chosen model

    Run speech recognition — defaults to if is omitted

  4. 4

    Your app

    Receive (plus in )

Tip

Leave model unset to get saaras:v4, and leave language unset to let it auto-detect. Set mode=translate to get English text out of any supported language in a single call.

Streaming-only models cannot transcribe files. ink-2, the Deepgram Flux models, assemblyai-universal-3.6-pro and inworld-stt-1 do turn detection as part of the model, which only makes sense on a live stream. Sending one here returns an error naming the file-transcription alternative rather than silently substituting a different model — see Cartesia Ink models.

Basic Usage

from openai import OpenAI

client = OpenAI(
    api_key="cm_your_key",
    base_url="https://api.callmissed.com/v1"
)

with open("audio.wav", "rb") as f:
    response = client.audio.transcriptions.create(
        model="saaras:v4",
        file=f
    )

print(response.text)

Parameters

Send as multipart/form-data.

ParameterTypeRequiredDescription
filefileYesAudio file (WAV, MP3, etc.), up to 25 MB. An empty file returns 400. Some models accept less per request: saaras:v4 takes at most 30 seconds of audio, gnani-prisma-v3.0 (and its retired ID gnani-prisma-v2.5) at most 30 seconds, and whisper, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize at most 25,000,000 bytes
modelstringNoDefault saaras:v4. Any file-transcription STT model ID from Models. ink-2, scribe_v2_realtime and the Flux models are not valid here — see Cartesia Ink models and ElevenLabs Scribe models
languagestringNoLanguage code; auto-detected if omitted. For saaras:v4 use a locale such as hi-IN or ta-IN — a code the model does not support falls back to auto-detection. For gnani-prisma-v3.0 use one of hi-IN, bn-IN, ta-IN, te-IN, kn-IN, ml-IN, mr-IN, gu-IN, pa-IN, en-IN, a comma-separated list of up to 4 of them, or indic-auto to detect the language; a base code such as hi is read as hi-IN
modestringNoOutput mode — see below. Default transcribe; any other value returns 422
response_formatstringNojson (default), text, verbose_json, or diarized_json. Any other value is answered as json
temperaturenumberNo0–2. Accepted for OpenAI SDK compatibility; currently not forwarded to the model
promptstringNoAccepted for OpenAI SDK compatibility; currently not forwarded to the model

Response

json (default):

{"text": "Namaste, aap kaise hain?"}

text returns the transcript as text/plain. verbose_json adds the billed audio duration in seconds. words is always an empty array — word-level timestamps are not returned. segments is empty except on gpt-4o-transcribe-diarize, where each segment carries id, speaker, start, end and text. diarized_json returns task, duration, text and those speaker segments.

{
  "task": "transcribe",
  "language": "hi-IN",
  "duration": 4.52,
  "text": "Namaste, aap kaise hain?",
  "segments": [],
  "words": []
}

language echoes the language you sent, or "unknown" when you let the model auto-detect.

Errors

Errors use the OpenAI envelope: {"error": {"message", "type", "code"}}.

StatuscodeWhen
400invalid_requestfile is empty; the audio or a parameter was rejected (format, size, language); or the model is streaming-only or not available for file transcription. The message says which
400audio_too_longThe audio is longer than the model accepts per request (see file above). Use a model without a duration limit, such as deepgram-nova-3
402insufficient_creditsCredit balance is exhausted
403permission_deniedThe API key does not have the stt permission, or the model is not available on your plan
404model_not_foundmodel is not a known STT model ID
422—A form field fails validation (for example an unknown mode, or temperature outside 0–2)
413file_too_largeThe file is larger than 25 MB, or than the model accepts (see file above)
429quota_exceededPlan usage limit reached
429rate_limit_exceededTranscription is rate limited right now. Retry with backoff
503service_unavailableTranscription is temporarily unavailable. Retry shortly. /v1/audio/speech uses the same code
503provider_error, upstream_timeout, upstream_unavailableThe model failed to transcribe the file. The response includes a request_id

/v1/audio/translations returns the same errors.

Output Modes

ModeDescription
transcribeStandard transcription (default)
translateTranscribe and translate to English
verbatimExact transcription including filler words
translitTransliteration to Latin script
codemixCode-mixed output (Indic + English)

mode is applied by saaras:v4. whisper also honours translate. Every other model ignores mode and returns a plain transcript.

saaras:v4 serves all five modes on one model across 23 languages, and is on the free tier:

curl -X POST https://api.callmissed.com/v1/audio/transcriptions \
  -H "Authorization: Bearer cm_your_key" \
  -F [email protected] \
  -F model=saaras:v4 \
  -F mode=codemix

Cartesia Ink models

Two Cartesia STT models, and they are not interchangeable — one transcribes files, the other only runs on a live voice session.

ModelPriceLanguagesFile transcriptionVoice sessions
ink-whisper$0.1875 / hr100 (incl. Hindi, Urdu, Tamil)YesYes
ink-2$0.5625 / hr5: en, fr, hi, ja, esNoYes

ink-whisper — the cheapest 100-language option

Cartesia's fastest and most affordable STT, with better accuracy than baseline Whisper. Its dynamic chunking cuts hallucination during pauses and silence, so audio with dead air transcribes cleanly instead of inventing text to fill gaps.

curl -X POST https://api.callmissed.com/v1/audio/transcriptions \
  -H "Authorization: Bearer cm_your_key" \
  -F [email protected] \
  -F model=ink-whisper \
  -F language=hi

ink-2 — voice agents only

Cartesia's top-ranked STT for voice agents: 8% WER on AppTek's 14-accent call-centre benchmark, against 10% for Deepgram Flux and 12% for ElevenLabs. It also self-detects turns, so a voice agent needs no separate turn detector on top.

Two limits decide whether you can use it at all:

1. It cannot transcribe files. ink-2 is streaming-only. POSTing it to /v1/audio/transcriptions fails rather than quietly substituting a different model. The request is not transcribed, and the error message names the alternative:

{
  "error": {
    "message": "ink-2 is a streaming-only model and is not available for file transcription. Use ink-whisper here, or ink-2 on a voice session.",
    "type": "invalid_request_error",
    "code": "invalid_request",
    "request_id": "stt-…"
  }
}

The HTTP status is 400. deepgram-flux-general-en and deepgram-flux-general-multi fail the same way and point you to deepgram-nova-3.

Select it on a voice session or the Managed Voice Agent instead.

2. It covers five languages. The model accepts en, fr, hi, ja and es. Sending it any other language (Tamil, for example) does not raise an upstream error — it silently produces poor output. Our agent logs a warning and transcribes as en. For other languages use ink-whisper (100 languages) or an Indic model such as saaras:v4.

ElevenLabs Scribe models

Three ElevenLabs speech-to-text models. Two transcribe files; the third only runs on a live voice session. All three hear the same 100 languages, including Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, Urdu and Sinhala, and detect the language automatically when language is left out. Paid plans only.

ModelPriceFile transcriptionVoice sessions
scribe_v2$0.2292 / hrYesNo
scribe_v2_medical$0.2292 / hrYesNo
scribe_v2_realtime$0.4063 / hrNoYes

scribe_v2 is ElevenLabs' state-of-the-art recognizer for recorded audio. scribe_v2_medical is the same API and price, fine-tuned for clinical audio (clinician-patient conversations, dictation, intake calls) with 35% fewer errors on it and the same accuracy on everyday speech. language takes a base code (hi; hi-IN is reduced to hi); a code the model does not hear is ignored and the language is detected instead. mode other than transcribe is not applied. Silence counts toward the billed duration.

scribe_v2_realtime is the streaming model for voice agents (~150 ms to a committed transcript upstream): select it with stt_model on an agent or a voice session. It has no turn detection of its own, so the agent's turn detection applies. POSTing it here returns 400; use scribe_v2 for files.

curl -X POST https://api.callmissed.com/v1/audio/transcriptions \
  -H "Authorization: Bearer cm_your_key" \
  -F "model=scribe_v2" \
  -F "language=hi" \
  -F "[email protected]"

Voice-session-only models

Four more speech-to-text models run only on a live voice session, selected with stt_model on an agent or a voice session. POSTing any of them to /v1/audio/transcriptions returns 400. For files, use saaras:v4, ink-whisper or deepgram-nova-3.

ModelPriceLanguagesTurn detection
assemblyai-universal-3.6-pro$0.4688 / hr50Built into the model
inworld-stt-1$0.1563 / hrEnglish only (en, en-US)Not built in
microsoft-speech-stt$1.0417 / hr81 (148 locales), one per callNot built in
scribe_v2_realtime$0.4063 / hr100, auto-detectedNot built in

assemblyai-universal-3.6-pro is AssemblyAI's newest streaming model. It ends the caller's turn itself, so the agent needs no separate turn detector on top, and it auto-detects and code-switches, so the language you set is a hint. Languages include English (with en-IN, en-US, en-GB, en-AU, en-CA, en-NZ), Spanish, French, German, Italian, Portuguese, Turkish, Dutch, Swedish, Norwegian, Danish, Finnish, Hindi, Marathi, Urdu, Vietnamese, Arabic, Hebrew, Japanese, Chinese, Cantonese, Russian, Korean, Catalan, Galician, Romanian, Estonian, Persian, Afrikaans, Zulu and Xhosa.

inworld-stt-1 is Inworld's low-cost streaming model and hears English only. For any other language use assemblyai-universal-3.6-pro, saaras:v4 or deepgram-nova-3.

scribe_v2_realtime is ElevenLabs' real-time Scribe model: 100 languages with automatic detection, so a multilingual agent can follow the caller. See ElevenLabs Scribe models.

microsoft-speech-stt is Microsoft's real-time streaming recognizer, processed in India. It hears 81 languages (148 locales), including Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, Urdu, Bhojpuri and Indian English. It recognises one language per call: it does not auto-detect or follow a caller who switches language, so set the agent's language to the one the caller speaks. It has no turn detection of its own, so the agent's turn detection applies. Keyterms you set on the agent (stt_keyterms) are applied as a phrase list.

mai-transcribe-2 (preview)

mai-transcribe-2 is Microsoft AI's MAI-Transcribe-2, $0.375 / hour. It is preview. It transcribes files on POST /v1/audio/transcriptions only: it is not available in voice sessions, and POST /v1/audio/translations refuses it.

  • Languages. 60: af, ar, as, az, bg, bn, bs, ca, cs, da, de, el, en, es, et, fa, fi, fil, fr, gl, gu, he, hi, hu, hy, id, is, it, ja, kk, kn, ko, lt, lv, mk, ml, mr, ms, nb, ne, nl, or, pa, pl, pt, ro, ru, sk, sl, sv, sw, ta, te, th, tr, uk, ur, vi, yue, zh.
  • Language detection. Leave language out and the language is detected automatically, including Hinglish and other code-switching. language takes one code; hi and hi-IN both work. A language not in the list returns 400.
  • Files. Up to 2 hours of audio per file and the route's 25 MB upload limit. WAV, MP3, FLAC and Ogg are accepted. response_format is json, text or verbose_json.
  • Plan and hosting. Paid plans only, never a default, and not processed in India.
  • Preview. Microsoft ships this model as a public preview with no SLA, and may change or retire it at short notice. If it is retired, the ID keeps working and is answered by a named successor, as described in Retired and deprecated models, so the transcript may differ. Use a generally available model for production-critical workloads.
curl -X POST https://api.callmissed.com/v1/audio/transcriptions \
  -H "Authorization: Bearer cm_your_key" \
  -F "model=mai-transcribe-2" \
  -F "[email protected]"

Deepgram feature parameters

When you select a Deepgram file-transcription model (deepgram-nova-3, deepgram-nova-2, deepgram-enhanced, etc.), these extra form fields are accepted. They are ignored for non-Deepgram models. Model-restricted features are dropped automatically when the chosen model doesn't support them.

ParameterTypeDescription
diarizebooleanLabel each speaker ([Speaker 0], [Speaker 1], …)
utterancesbooleanSegment the transcript into utterances
utt_splitnumberSilence gap (seconds, 0–10) used to split utterances
paragraphsbooleanSplit the transcript into paragraphs
numeralsbooleanWrite numbers as digits (e.g. "five" → "5")
measurementsbooleanAbbreviate measurement units (English)
dictationbooleanConvert spoken "comma"/"period" to punctuation (English)
profanity_filterbooleanMask recognized profanity with ****
filler_wordsbooleanKeep "uh"/"um" (Nova / Nova-2 / Nova-3)
multichannelbooleanTranscribe each audio channel independently
detect_entitiesbooleanTag entities like names and locations (English)
detect_languagestringtrue to auto-detect, or repeat with codes to restrict
redactstringpci, pii, phi, numbers, or a specific entity type (repeatable)
keytermstringBoost recognition of a term/phrase (Nova-3, including medical and pharma; repeatable)
language_hintstringSteer language=multi toward the languages you expect, e.g. language_hint=hi and language_hint=en. deepgram-nova-3 only, and only with language=multi; ignored otherwise, as are codes Deepgram does not list. Repeatable
keywordsstringkeyword:intensifier boost/suppress (Nova-2 / legacy; repeatable)
searchstringPhonetically search the audio for a term (repeatable)
replacestringfind:replacement substitution (repeatable)

Dialects & locales

Deepgram models accept locale-specific language codes so you can pin a dialect for best accuracy. Pass the code in the language field. Each model's exact dialect list is published in the dialects array on GET /v1/models. Examples:

  • English: en-US, en-GB, en-IN, en-AU, en-NZ, en-CA, en-IE
  • Spanish: es, es-419 (Latin America)
  • Portuguese: pt-BR, pt-PT
  • Chinese: zh-CN, zh-TW, zh-HK (Cantonese)
  • Multilingual code-switching: multi (Nova-3, Nova-3 Medical, Nova-2). Nova-3 Medical covers 10 languages in multi; deepgram-nova-3-pharma is English only

On this page