Transcribe non-English or mixed-language speech in the language spoken
Four or more companies' docs show a full fit.
What counts as done
Text in the spoken language (not a translation), with the language either forced by a parameter or detected automatically, for every language the listing claims.
Measured, for our own tests: where the language is detected rather than forced, it is detected correctly on a 10-second clip. No docs we read state these figures, so they don't count against a listing here.
How agents ask for it: transcribe this French interview, in French
, this call switches between Hindi and English
, transcribe this Irish recording
Does the job 15
The docs show it meets the test above, and nothing we read says otherwise.
| Listing | Why |
|---|---|
| Amazon Transcribe BB Amazon Web Services, Inc. | LanguageCode forces the language; IdentifyLanguage and IdentifyMultipleLanguages detect one or several languages in batch and streaming, with transcription in the spoken language. source |
| AssemblyAI Speech-to-Text (Universal) B AssemblyAI, Inc. | language_code forces the language; language_detection and code switching on Universal-3.6 Pro (32 languages live, 99 on Universal-2). Irish is not covered (acc-14). source |
| Azure AI Speech speech-to-text BB Microsoft Corporation | Locale forces the language; language identification takes up to 4 candidates at start or 10 continuous, and batch languageIdentification lists candidateLocales. source |
| Deepgram Speech-to-Text (Nova-3, Flux) BB Deepgram, Inc. | language forces the language, detect_language picks one, and language=multi code-switches across 55 Nova-3 languages; Flux Multilingual covers 10 with language_hint. Irish is absent (acc-14). source |
| ElevenLabs Scribe Speech to Text API BB Eleven Labs Inc. | language_code forces the language, secondary_language_codes declares others on scribe_v3, detection is automatic when omitted, and the capability page lists Irish (gle). source |
| Gladia Speech-to-Text API + MCP BB Gladia SAS | language_config.languages forces one language or lists several; code_switching detects per utterance on solaria-1 across 100+ languages. solaria-3 is single-language. source |
| Google Cloud Speech-to-Text B Google LLC | language_codes forces one or several languages, Chirp 3 infers the prevalent language, and multi-language recognition is documented; acc-29 saw 12 percent misidentification on short clips. source |
| Groq Speech-to-Text BB Groq LLC | language forces one of Whisper's languages and detection runs when omitted; no code-switching handling is documented. source |
| MiniMax Speech-to-Text D Nanonoble Pte. Ltd. | A BCP-47 language header forces the language; omitting it enables mixed-language recognition across the 20 listed languages. source |
| Nemotron 3.5 ASR Streaming 0.6B C NVIDIA Corporation | language or target_lang takes a locale or bare code, or auto, which emits a detected language tag; 32 locales transcribe out of the box and 8 need fine-tuning. source |
| OpenAI Speech to Text BB OpenAI OpCo, LLC | languages lists expected languages and the response reports detected languages; transcription is in the spoken language, translation being a separate endpoint. source |
| Qwen3-ASR-0.6B C Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check) | language forces one of 30 languages and 22 dialects or None detects; the card lists identification and recognition for both. source |
| Qwen3-ASR-1.7B B Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check) | language forces one of 30 languages and 22 dialects or None detects; the card lists identification and recognition for both. source |
| Soniox Speech-to-Text B Soniox Inc. | Auto-detects 60+ languages with mid-sentence switching; language_hints bias and language_hints_strict restricts, so the language can be forced. source |
| Speechmatics Speech-to-Text BB Cantab Research Ltd | language forces one of 55+ languages, auto runs language identification, bilingual packs cover mixed audio, and Melia 1 switches languages automatically. source |
Does the job, with a limit 12
It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.
| Listing | Why |
|---|---|
| Cartesia Ink B Cartesia AI, Inc. | language forces ink-whisper across about 43 languages; ink-2 covers en, fr, hi, ja and es only and the language parameter does not apply to it. source |
| Cohere Transcribe 03-2026 C Cohere Inc. | language= must be given from 14 codes; there is no detection, and the blog says code-switching was not trained for. source |
| Giggy Speech to Text E Giggy | language forces one of 27 listed languages or detection runs by default; no code-switching is documented. source |
| IBM Watson Speech to Text C International Business Machines Corporation | model selects the language; /v1/detect_language identifies a spoken language, but no multi-language or code-switching mode is documented. source |
| Inworld STT C Theai, Inc. dba Inworld AI | language is a hint that biases automatic detection across 30 languages, not a hard force. source |
| Ming-3.0-asr-tiny C inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI) | --language-hint zh or en only; the model also labels the detected language before a tab, but covers Chinese, English and dialects. source |
| Ming-3.0-audio-flash E inclusionAI, Ant Group (named in the inclusion-ai.org footer; no registered company named) | Chinese and English only; the model labels the language before a tab rather than taking a forced-language parameter. source |
| Ming-3.0-audio-tiny D inclusionAI (Ant Group) | Chinese and English only, with a language label before a tab rather than a forcing parameter. source |
| Mistral Voxtral Transcribe B Mistral AI (RCS Paris 952 418 325) | language forces one of 13 languages and detection runs when omitted, but setting language disables timestamps. source |
| Rev AI Speech-to-Text API C Rev.com, Inc. | language selects one language per job or stream; no detection or code-switching documented, and streaming speaker switch is English and Spanish only. source |
| Smallest.ai Pulse STT B Smallest, Inc. | language must be set on every call; 21 codes on streaming and 12 on pre-recorded, with regional auto-detect aggregators rather than general detection. source |
| Voxtral Mini 4B Realtime 2602 B Mistral AI (RCS Paris 952 418 325) | 13 languages with the model detecting the language; no documented parameter forces it. source |
Doesn't do this job 2
Listings in the category that do something else, with the reason.
| Listing | Why |
|---|---|
| Parakeet TDT 0.6B v3 B NVIDIA Corporation | No language parameter; acc-07 and acc-24 show French audio returned as English translation and acc-23 shows mixed-language drift. |
| Underdog Bark 3.0 E Conway Research | English only. |
3 listings in the category were added after this reading and aren't marked yet.
What stops agents finishing
- model translates into English instead of transcribing (acc-07, acc-24 on Parakeet v3)
- no way to force a language (Parakeet v3) or no detection at all (Cohere Transcribe)
- language not covered (Irish in acc-14; Ming, Voxtral and Cartesia Ink 2 cover 2 to 13 languages)
- short clips misidentified about 12 percent of the time (acc-29)
Around it
Before: language.identify. After: translation.translate-text, llm.summarise.
Where this comes from
Weight 10 of 100 in Speech-to-text: 6 of 32 accounts (acc-07, acc-14, acc-18, acc-23, acc-24, acc-29); on open weights the complaint was drift and translation, on APIs it was coverage
The same as data: /api/v1/use-cases/speech-to-text.multilingual-known-language.json, or the MCP tool get_use_case with id speech-to-text.multilingual-known-language. Other jobs in this category: /api/v1/niches?category=speech-to-text.
For companies
Do agents find, use and choose your tools?
An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.