# Best speech to text APIs for multilingual and mixed-language audio

Which transcription API fits audio in Thai, English, Mandarin, or two languages in one sentence, with language detection and live use, from each vendor's own docs.

Published 2026-11-17. Paxa Labs.

For audio that mixes languages inside a sentence, the documentation of five services states the fit. Paxa STT Lite writes each word as it was spoken with no language setting. Soniox, Speechmatics Melia 1, Gladia Solaria-1, and Amazon Transcribe each document switching between languages in one recording. For a single language, start from the vendor's language list and read it model by model.

Every vendor fact below links that vendor's own documentation, as our comparison pages verified it in October and November 2026. No vendor is scored or ranked. Published rates for each one are on [the speech to text API comparison](https://paxalabs.com/compare/speech-to-text-apis), re-checked every month.

> Paxa STT Lite and Paxa STT Lite Realtime are both in Research preview. Their transcripts and options may change between releases, and [the speech to text page](https://paxalabs.com/speech-to-text#changelog) records each change.

## Multilingual speech to text APIs at a glance

| Service | Languages | Mixed languages and detection | Subtitles and word timing | Live audio |
| --- | --- | --- | --- | --- |
| Paxa STT Lite (Research preview) | Multilingual, with Thai, English, and Mandarin named | Writes mixed speech as spoken and detects the language, with an optional hint | Word timestamps, SRT, and WebVTT in the same response | Paxa STT Lite Realtime (Research preview), with automatic turn detection |
| Soniox | 60+ languages at one price | Follows a change of language within a sentence, and each token can carry its language | Start and end time on every token | WebSocket, with an end token when the speaker finishes |
| Speechmatics | A published list that includes Thai | Melia 1 switches between the listed languages with no selection | Word timings on batch | Melia 1 realtime, where the client marks the end of an utterance |
| Gladia | Solaria-1 covers every language on its table, Thai and Chinese included | A code switching setting returns the detected language with each utterance and word | SRT and VTT export | Live sessions on Solaria-1 |
| Amazon Transcribe | Thai and Simplified Chinese listed for batch and streaming | Multi-language identification covers speakers who change language mid-conversation | WebVTT and SubRip from batch video jobs | Streaming over HTTP/2 or WebSockets |
| AssemblyAI | 18 languages on Universal-3.5 Pro, 99 on Universal-2 | Universal-3.5 Pro follows speakers mid-sentence across its 18 languages | SRT and VTT export | Streaming with a message for each finished turn |
| Deepgram | 45+ languages on Nova models | Nova-3's multilingual option covers ten languages | A captions package that builds WebVTT or SRT in your code | Flux, in English or across 10 languages |
| ElevenLabs Scribe | 90+ languages on Scribe v2 | Predicts the language when no language code is sent | Absent from the pages cited here | Realtime, with manual or silence-based commits |
| Google Chirp 3 | Thai listed at Preview launch readiness | Set to auto, it transcribes the dominant language of the file | Word timestamps on batch audio up to 20 minutes | Absent from the page cited here |
| Rev AI | 58+ languages on files, 9+ on streaming | One multilingual entry, English with Spanish | SRT or VTT captions | Streaming in 9+ languages, and Thai is files only |

## Which transcription API is best for your language?

The answer changes with the language, because a vendor's headline count hides which model reads which language. Check the list for the exact model you plan to call.

Thai is on the lists of [Deepgram](https://developers.deepgram.com/docs/models-languages-overview) for Nova-3 and Nova-2, [Speechmatics](https://docs.speechmatics.com/speech-to-text/languages), [ElevenLabs](https://elevenlabs.io/docs/overview/capabilities/speech-to-text) for Scribe v2, [Gladia](https://docs.gladia.io/chapters/language/supported-languages) for Solaria-1, and [Amazon Transcribe](https://docs.aws.amazon.com/transcribe/latest/dg/supported-languages.html) for batch and streaming. [Google's Chirp 3 page](https://docs.cloud.google.com/speech-to-text/v2/docs/chirp_3-model) lists Thai at Preview launch readiness. [AssemblyAI](https://www.assemblyai.com/docs/pre-recorded-audio/supported-languages) lists Thai for Universal-2 and leaves it off Universal-3.5 Pro. [Rev AI](https://www.rev.ai/languages) lists Thai for files with no streaming. [The Thai speech to text guide](https://paxalabs.com/guides/best-thai-speech-to-text) goes through Thai job by job.

Mandarin is among the 18 languages of [AssemblyAI's Universal-3.5 Pro](https://www.assemblyai.com/docs/pre-recorded-audio/supported-languages). [Deepgram](https://developers.deepgram.com/docs/models-languages-overview) lists Chinese (Mandarin, Simplified) for Nova-3 and Nova-2, [Amazon Transcribe](https://docs.aws.amazon.com/transcribe/latest/dg/supported-languages.html) lists zh-CN for batch and streaming, [Gladia](https://docs.gladia.io/chapters/language/supported-languages) lists Chinese on Solaria-1, and [Rev AI](https://www.rev.ai/languages) lists Mandarin Chinese (Simplified).

For European languages, [Gladia](https://docs.gladia.io/chapters/language/supported-languages) marks Solaria-3 for English, French, German, Italian, and Spanish. [Deepgram's multilingual option](https://developers.deepgram.com/docs/models-languages-overview) covers English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch.

Paxa STT Lite, a Research preview, is multilingual and transcribes Thai, English, Mandarin, and other languages. A team whose audio is all in one of the European languages above has more documented choices among the other vendors.

## What is the best speech to text API for bilingual or mixed-language speech?

Bilingual speakers change language inside a sentence. A Thai engineer says a product name in English and finishes the thought in Thai. A useful transcript keeps each word in the script it was spoken in.

[Soniox's Thai page](https://soniox.com/speech-to-text/thai) says its model follows changes between Thai and other languages within the same sentence. On [Speechmatics](https://docs.speechmatics.com/speech-to-text/languages), Melia 1 switches between the listed languages with no language selection, while Enhanced and Standard take one selected language or a fixed pack. [Gladia's code switching guide](https://docs.gladia.io/chapters/language/code-switching) turns switching on with a setting and recommends a short list of expected languages. [Amazon Transcribe's multi-language identification](https://docs.aws.amazon.com/transcribe/latest/dg/lang-id-batch.html) transcribes each supported language in a file, including when speakers change languages mid-conversation.

Two vendors limit the mix. [AssemblyAI's Universal-3.5 Pro](https://www.assemblyai.com/docs/pre-recorded-audio/code-switching) follows speakers mid-sentence across its 18 languages, and Universal-2 detects up to two languages per file, one of them English on a manual list. [Rev AI](https://www.rev.ai/languages) has one multilingual entry, English with Spanish.

Paxa STT Lite, a Research preview, writes a sentence that switches languages as it was spoken, with no language setting and no translation. [The code-switching guide](https://paxalabs.com/guides/transcribing-thai-english-code-switching) shows a Thai and English recording and what the transcript does for search.

## Which speech to text APIs detect the spoken language automatically?

Detection matters when recordings arrive unlabeled, as in a support queue that serves several countries. [ElevenLabs' API reference](https://elevenlabs.io/docs/api-reference/speech-to-text/convert) predicts the language when no `language_code` is sent. [Soniox](https://soniox.com/docs/stt/async/async-transcription) puts a language field on each token when language identification is on, and [its pricing page](https://soniox.com/pricing) includes identification in the hourly rate. [Gladia](https://docs.gladia.io/chapters/language/supported-languages) marks auto-discovery for Thai and Chinese on Solaria-1.

[Google's Chirp 3](https://docs.cloud.google.com/speech-to-text/v2/docs/chirp_3-model) set to auto transcribes the dominant language of a file. [Amazon Transcribe](https://docs.aws.amazon.com/transcribe/latest/dg/lang-id-batch.html) names en-US, th-TH, and zh-CN among the languages of its multi-language identification, and states that redaction and custom language models are off while it runs.

Paxa STT Lite, a Research preview, detects the language when the request names none. A language hint is optional, and it changes no charge.

## Which speech to text APIs come with translation built in?

Three vendors list translation beside transcription. [Speechmatics' pricing page](https://www.speechmatics.com/pricing) includes translation among its add-ons for recorded audio. [Gladia](https://www.gladia.io/pricing) prices translation as an hourly add-on, and [Soniox](https://soniox.com/pricing) bills translation by token usage.

Paxa keeps the two steps separate. Paxa STT Lite returns the transcript in the language spoken, and [the translation API](https://paxalabs.com/translation) on the same account turns it into another language, with a glossary for product names. [How to dub a video with AI](https://paxalabs.com/guides/how-to-dub-a-video-with-ai) chains the two with a voice.

## Which multilingual APIs transcribe live audio?

A live transcript has to decide where each turn ends. [Soniox's endpoint detection](https://soniox.com/docs/stt/rt/endpoint-detection) returns an `<end>` token once its model judges the speaker finished. [AssemblyAI's streaming API](https://www.assemblyai.com/docs/streaming/getting-started/transcribe-streaming-audio) sends a Turn message with `end_of_turn` set to true. [Deepgram's Flux](https://developers.deepgram.com/docs/flux/quickstart) detects the end of a turn inside the model, in English or across 10 languages with flux-general-multi. On [Speechmatics' Melia 1](https://docs.speechmatics.com/speech-to-text/models), the client sends ForceEndOfUtterance when the speaker has finished.

Live language coverage is often narrower than file coverage. [Rev AI's FAQ](https://docs.rev.ai/faq) states 58+ languages on files and 9+ on streaming. [Gladia's live guide](https://docs.gladia.io/chapters/live-stt/getting-started) runs live transcription on Solaria-1 only. With [OpenAI's gpt-live-transcribe](https://developers.openai.com/api/docs/guides/realtime-transcription), the application commits each turn after detecting the end of speech on its side.

Paxa STT Lite Realtime, a Research preview, is multilingual streaming transcription with automatic turn detection. It reports when the user starts and stops talking and sends one final transcript per turn. [The realtime speech to text page](https://paxalabs.com/realtime-speech-to-text) shows a session, and [realtime against recorded transcription](https://paxalabs.com/guides/realtime-vs-recorded-transcription) helps choose between the two.

## Which API fits captions, subtitles, and media monitoring?

Captioning needs a subtitle file, and monitoring needs long recordings in several languages read on a schedule. [AssemblyAI](https://www.assemblyai.com/docs/pre-recorded-audio/export-transcripts-as-srt-vtt-or-text), [Gladia](https://docs.gladia.io/chapters/audio-intelligence/subtitles), and [Rev AI](https://docs.rev.ai/api/asynchronous/reference/) each export SRT and VTT. [Amazon Transcribe](https://docs.aws.amazon.com/transcribe/latest/dg/subtitles.html) writes WebVTT and SubRip files for batch video jobs into the same S3 location as the transcript.

Broadcast recordings run long. [Rev AI](https://docs.rev.ai/faq) states a maximum audio length of 17 hours, and [ElevenLabs Scribe](https://elevenlabs.io/docs/overview/capabilities/speech-to-text) takes files up to 10 hours in standard mode. [Amazon Transcribe](https://docs.aws.amazon.com/general/latest/gr/transcribe.html) sets a maximum of 28,800 seconds and 2 GB per file. [Gladia](https://docs.gladia.io/chapters/limits-and-specifications/supported-formats) accepts 135 minutes per request, and longer on enterprise plans.

Paxa STT Lite, a Research preview, transcribes up to 120 minutes and 95 MiB in one request and returns SRT or WebVTT in the same response, with every line break on a word boundary. Split a longer broadcast at a pause. [The subtitle generator page](https://paxalabs.com/subtitle-generator), [the captions page](https://paxalabs.com/use-cases/captions), and [the media monitoring page](https://paxalabs.com/use-cases/media-monitoring) show each workflow, and [the word-level timestamps guide](https://paxalabs.com/guides/word-level-timestamps-api) explains the timing behind a caption.

## Which multilingual API is most accurate across languages?

No public figure answers that for your audio. Accuracy moves with the language, the microphone, the accent, and how often speakers switch. A vendor that leads on English podcasts can trail on a Thai call recording. The reliable method is a test on your own recordings.

- Pick three recordings like your real traffic, such as a clean one, a noisy one, and one where two languages mix.
- Run each through every vendor listed under your question, with the same settings.
- Read the transcripts against the audio at names, figures, and the words where the language changes.
- Score Thai and Chinese on characters. [The accuracy guide](https://paxalabs.com/guides/measuring-thai-transcription-accuracy) explains why word error rate misleads on a language written without spaces.

## What should you read or try next?

- Upload a mixed-language recording to [the Transcribe tool](https://paxalabs.com/app/transcribe). A new account's 100 free credits cover about 12 minutes of audio, with no card.
- Read what the models take and return on [the speech to text page](https://paxalabs.com/speech-to-text) and [the realtime speech to text page](https://paxalabs.com/realtime-speech-to-text).
- See a recording transcribed word by word on [the online transcription page](https://paxalabs.com/transcribe).
- Compare published rates on [the speech to text API comparison](https://paxalabs.com/compare/speech-to-text-apis).
- New to the subject? Start with [what speech to text is](https://paxalabs.com/guides/what-is-speech-to-text), then see [what transcription costs](https://paxalabs.com/guides/how-much-does-transcription-cost).

## Common questions

### Is Amazon Transcribe a suitable option for converting Thai video audio into text?

It can do the job. [Amazon Transcribe's language table](https://docs.aws.amazon.com/transcribe/latest/dg/supported-languages.html) lists Thai (th-TH) for batch and streaming, and [its input page](https://docs.aws.amazon.com/transcribe/latest/dg/how-input.html) reads MP4 and WebM files from an Amazon S3 bucket. Batch video jobs can return WebVTT and SubRip subtitles. The same table reads no for Thai under custom language models and redaction. [The Amazon Transcribe comparison](https://paxalabs.com/compare/aws-transcribe) sets it beside Paxa STT Lite.

### Which speech to text API providers serve European buyers?

Three facts from vendor pages help here. [Gladia](https://www.gladia.io/pricing) publishes its rates in euros. [Soniox's Thai page](https://soniox.com/speech-to-text/thai) lists an EU processing region beside the US, Japan, and India. [Speechmatics' Enterprise plan](https://www.speechmatics.com/pricing) lists on-premises containers and on-device deployment. Confirm the region and the contract terms with the vendor before you send audio.

### Can one API transcribe every language in my product?

Often, but check the model and the mode. A language listed for files may be missing from streaming, and a newer model may cover fewer languages than the one before it. A product can also send different languages to different vendors, each with its own key.

### Do multilingual transcripts label who spoke?

Most services in the table offer speaker labels on recorded audio. Paxa STT Lite, a Research preview, returns numbered speaker turns on a recording up to 120 minutes and identifies nobody by name. [The speaker turns guide](https://paxalabs.com/guides/thai-call-transcripts-with-speaker-turns) compares how vendors label speakers.

### What does Research preview mean for Paxa STT Lite?

Paxa STT Lite and Paxa STT Lite Realtime are live on the API and in the console while they are evaluated. Their transcripts and options may change between releases, and the changelog on [the speech to text page](https://paxalabs.com/speech-to-text#changelog) records each change.

### Do you publish accuracy comparisons?

No. We publish no accuracy, quality, or latency figures, for our models or anyone else's. Transcribe the same recording on each service you are considering and read the transcripts side by side.
