In short: AssemblyAI transcribes 99 languages from one API. AssemblyAI-STT-3.7 code-switches natively across 32 languages in recorded audio, and AssemblyAI-STT-3.7-Realtime does the same across 32 languages in real time. Both detect the spoken language, label speakers, and accept prompts and key terms. For recorded audio in languages outside the 32, add Universal-2 to your request.
Most speech-to-text APIs were built for one language at a time. Real audio doesn't work that way. A support call switches between English and Spanish mid-sentence. A conference panel has speakers from six countries. A doctor dictates in German and drops in English drug names. If your pipeline can't handle that, you're either picking the wrong model per file or losing accuracy where it matters most.
This guide covers what multilingual transcription means, which languages each model supports, and how to detect, diarize and transcribe mixed-language audio in code.
What is multilingual transcription?
Multilingual transcription is automatic speech recognition that works across more than one language, including when several languages appear in the same recording. Three separate capabilities get bundled under that one term:
- Multi-language coverage. The model transcribes audio in any language it supports.
- Automatic language detection. You don't have to tell the API what language the audio is in.
- Code switching. A speaker moves between languages within the same utterance, and the model keeps up without dropping words.
Coverage is table stakes now. Detection is common. Code switching is where models separate, and it's the capability teams underestimate until their real-world accuracy comes back lower than the demo. There's a second failure mode that gets less attention: a model that flips accented English into another language's script when the speaker never switched. A good multilingual model has to follow every real switch and leave everything else alone.
Supported languages: the full picture
Here's how coverage breaks down across AssemblyAI's models as of November 17, 2026.
| Model | Mode | Languages | Code switching | Price |
|---|---|---|---|---|
| AssemblyAI-STT-3.7 | Pre-recorded | 32 | Native, mid-sentence, among any languages you allow | $0.21/hr |
| Universal-2 | Pre-recorded | 99 | Opt-in, two languages, one must be English | $0.15/hr |
| AssemblyAI-STT-3.7-Realtime | Streaming: meetings, captions, calls | 32 | Native, on by default | $0.30/hr |
| AssemblyAI-STT-3.7-Realtime-Agent | Streaming: voice agents | 32 | Native, on by default | $0.45/hr |
Universal-3.5 Pro, the previous pre-recorded model with 18 languages, was deprecated on November 17, 2026. Starting December 8, 2026, existing accounts that don't set speech_models and default to Universal-3.5 Pro or Universal-3 Pro move to AssemblyAI-STT-3.7 automatically. Universal-Streaming Multilingual is also deprecated, with a tentative retirement date of February 1, 2027; AssemblyAI-STT-3.7-Realtime replaces it and serves all 32 languages in one stream. See model versions and deprecations.
AssemblyAI-STT-3.7 (pre-recorded): 32 languages with native code switching
AssemblyAI-STT-3.7 nearly doubles pre-recorded coverage, from 18 languages to 32:
- Carried over from Universal-3.5 Pro (18): English, Spanish, French, German, Italian, Portuguese, Arabic, Danish, Dutch, Finnish, Hebrew, Hindi, Japanese, Mandarin, Norwegian, Swedish, Turkish and Vietnamese.
- New (14): Urdu, Russian, Korean, Catalan, Galician, Romanian, Estonian, Persian, Cantonese, Afrikaans, Marathi, Zulu, Xhosa and Norwegian Nynorsk.
You don't need a dialect code. Pass the base code, such as es or fr, and the model handles regional variants like Mexican Spanish or Quebecois French. The supported languages docs list the dialects.
All 17 non-English languages that Universal-3.5 Pro supported get more accurate on AssemblyAI-STT-3.7, and their mean error falls from 12.8% to 10.1%:
| Language | Universal-3.5 Pro | AssemblyAI-STT-3.7 |
|---|---|---|
| Spanish | 8.5% | 6.9% |
| Mandarin* | 9.9% | 7.0% |
| Danish | 10.0% | 7.2% |
| Vietnamese | 10.3% | 7.4% |
| Finnish | 11.4% | 7.6% |
| Swedish | 8.3% | 7.7% |
| Portuguese | 9.2% | 7.8% |
| Italian | 9.2% | 8.8% |
| Norwegian | 11.6% | 9.3% |
| German | 10.9% | 9.8% |
| Dutch | 11.2% | 10.6% |
| French | 12.6% | 10.7% |
| Hebrew | 13.7% | 11.1% |
| Turkish | 14.4% | 11.5% |
| Japanese* | 16.2% | 14.1% |
| Hindi | 27.4% | 16.2% |
| Arabic | 22.9% | 18.1% |
Normalized word error rate, lower is better; * character error rate. Both models were run with the language specified in the request. Each figure is the mean over the benchmark datasets available for that language, covering read speech and conversational recordings. Source: the AssemblyAI-STT-3.7 research post.
Several of the new languages already rank among the model's most accurate: Galician at 4.1% word error rate, Russian at 4.6% and Catalan at 5.1%.
AssemblyAI-STT-3.7-Realtime (streaming): 32 languages with native code switching
All streaming models code-switch across the same 32 languages, with no configuration:
| Language | Code | Language | Code |
|---|---|---|---|
| Afrikaans | af | Korean | ko |
| Arabic | ar | Mandarin | zh |
| Cantonese | yue | Marathi | mr |
| Catalan | ca | Norwegian | no |
| Danish | da | Norwegian Nynorsk | nn |
| Dutch | nl | Persian | fa |
| English | en | Portuguese | pt |
| Estonian | et | Romanian | ro |
| Finnish | fi | Russian | ru |
| French | fr | Spanish | es |
| Galician | gl | Swedish | sv |
| German | de | Turkish | tr |
| Hebrew | he | Urdu | ur |
| Hindi | hi | Vietnamese | vi |
| Italian | it | Xhosa | xh |
| Japanese | ja | Zulu | zu |
The full list lives in the streaming multilingual docs. If you kept a language router in front of two streaming integrations, one model now covers both jobs.
Universal-2: 99 languages
For languages outside the 32, such as Polish, Ukrainian, Greek, Czech, Indonesian, Tamil and Thai, pre-recorded requests can run on Universal-2, which covers 99 languages. You add it to the request yourself, as the next section shows.
Sign up free and run AssemblyAI-STT-3.7 on your multilingual audio, with Universal-2 covering the rest of the 99 languages from the same API.
How to transcribe multilingual pre-recorded audio
Set speech_models to AssemblyAI-STT-3.7, turn on language detection, and the model follows speakers as they switch languages, transcribing every word in the language it was spoken. The examples use the AssemblyAI Python SDK; full scripts are in the companion repo.
import os
from assemblyai.prerecorded.v2 import Transcriber, TranscriptionConfig
config = TranscriptionConfig(
speech_models=["assemblyai-stt-3-7", "universal-2"], # Universal-2 covers languages outside the 32
language_detection=True,
speaker_labels=True,
)
transcript = Transcriber(api_key=os.environ["ASSEMBLYAI_API_KEY"], config=config).transcribe(
"https://assembly.ai/code-switching-3"
)
if transcript.status == "error":
raise RuntimeError(f"Transcription failed: {transcript.error}")
print(f"Model used: {transcript.speech_model_used}")
for utterance in transcript.utterances:
print(f"Speaker {utterance.speaker}: {utterance.text}")The speech_model_used field tells you which model handled the file. If you call the REST API directly, pass your raw API key in the authorization header, with no Bearer prefix.
Fallback rules on AssemblyAI-STT-3.7
Fallback works differently than it did on Universal-3.5 Pro, so check these before you migrate (from the migration guide):
language_detection: truewith onlyassemblyai-stt-3-7in the list: AssemblyAI always transcribes with AssemblyAI-STT-3.7 and never falls back. Add"universal-2"to the list to cover all 99 languages.language_codeset to a language AssemblyAI-STT-3.7 doesn't support: the request falls back to Universal-2.- Order: AssemblyAI-STT-3.7 takes priority wherever it appears in
speech_models, so["universal-2", "assemblyai-stt-3-7"]still sends every supported language to AssemblyAI-STT-3.7. - Verbatim by default: when
assemblyai-stt-3-7is anywhere in the list,disfluenciesdefaults totruefor the whole request, keeping fillers, repetitions and false starts. Set it tofalsefor clean transcripts. - Code-switching options:
code_switchingandcode_switching_confidence_thresholdinlanguage_detection_optionsapply only to Universal-2. AssemblyAI-STT-3.7 code-switches natively.
Choose your languages in code
When you already know which languages a file contains, pass language_codes to constrain the model to that set:
config = TranscriptionConfig(
speech_models=["assemblyai-stt-3-7"],
language_codes=["es", "fr"], # Spanish and French only; English isn't required
)- On AssemblyAI-STT-3.7,
language_codesaccepts any number of supported languages, and the model code-switches only among them. - A one-element list such as
["es"]pins the file to that language, the same aslanguage_code. language_codescan't be combined withlanguage_codeorlanguage_detection. Set one of the three.- If your list includes a language AssemblyAI-STT-3.7 doesn't support, the request runs automatic language detection and falls through to the next model in
speech_modelswhen needed. - On Universal-2,
language_codesaccepts at most two languages, and one must be"en".
Language detection works best with at least 15–90 seconds of speech. See the code switching docs for the full behavior, and this hands-on look at code switching and contextual prompting for examples on real audio.
How to transcribe multilingual audio in real time
AssemblyAI-STT-3.7-Realtime is multilingual by default. Open a streaming session with no language parameters and it code-switches across all 32 languages. Add language_detection and each turn reports which language was spoken, with a confidence score.
import os
from assemblyai.streaming.v3 import (
RealTimeEvents,
RealTimeParameters,
RealTimeTranscriber,
RealTimeTranscriberOptions,
TurnEvent,
)
def on_turn(client: RealTimeTranscriber, event: TurnEvent):
if event.end_of_turn:
print(f"[{event.language_code}] {event.transcript}")
client = RealTimeTranscriber(RealTimeTranscriberOptions(), api_key=os.environ["ASSEMBLYAI_API_KEY"])
client.on(RealTimeEvents.Turn, on_turn)
client.connect(
RealTimeParameters(
sample_rate=16000,
speech_model="assemblyai-stt-3-7-realtime",
language_detection=True,
# language_codes=["en", "es"], # optional: steer toward known languages
)
)
# Stream 16 kHz mono PCM with client.stream(chunk), then:
# client.disconnect(terminate=True)For a voice agent that answers the caller, use assemblyai-stt-3-7-realtime-agent instead; the language parameters are the same.
You can change the language set mid-stream without reconnecting. That's useful when a voice agent detects the caller's language or hands off to a speaker of another language. The change takes effect from the next turn:
from assemblyai.streaming.v3 import RealTimeSessionParameters
client.set_params(RealTimeSessionParameters(language_codes=["en", "fr"])) # steer
client.set_params(RealTimeSessionParameters(language_codes=[])) # back to full code switchingCode-switching accuracy: how the models compare
On pre-recorded audio that mixes English with German, Spanish, French, Italian or Portuguese (16 datasets), AssemblyAI-STT-3.7 has the lowest average error rate of the systems measured, and the lowest error rate on 12 of the 16 datasets, ahead of ElevenLabs Scribe v3, Soniox, Azure, AWS Transcribe, OpenAI, Deepgram and Gladia.
| Pre-recorded API | Average WER on code-switched audio (lower is better) |
|---|---|
| AssemblyAI-STT-3.7 | 6.2% |
| ElevenLabs Scribe v3 | 6.8% |
| Soniox stt-async-v5 | 10.2% |
| Azure Batch Transcription | 17.4% |
| Voxtral Mini Transcribe 2 | 17.4% |
| AWS Transcribe | 22.9% |
| OpenAI GPT-4o Transcribe | 44.7% |
| Deepgram Nova-3 | 49.4% |
| Gladia Solaria-1 | 50.4% |
Selected providers from the AssemblyAI-STT-3.7 research post, each vendor's production pre-recorded API at default settings. AssemblyAI-STT-3.7 and ElevenLabs Scribe v3 measured November 2026; other providers September 2026. Speechmatics, Gemini and Grok are left out because their APIs failed repeatedly on these datasets.
General-purpose transcription and purpose-built multilingual ASR aren't the same product, and the spread in that table shows it. The other half of the problem is staying put: across AssemblyAI's English benchmarks, stray characters from other scripts fell by 85% on AssemblyAI-STT-3.7, from 161 to 24 per million characters, so accented English stays in English.
In streaming, AssemblyAI-STT-3.7-Realtime posts 6.67% word error rate on code-switched audio (English paired with Spanish, German, French, Portuguese and Italian), first on all five language pairs among the 24 realtime models measured, ahead of ElevenLabs Scribe v3 Live, AWS Transcribe Streaming, Soniox Realtime v6 and Deepgram (per the AssemblyAI-STT-3.7 launch post). Comparing providers for your own audio? Our companion piece, multilingual speech-to-text in 2026, covers how to run that evaluation, and how to evaluate speech recognition models covers the method. The benchmarks page has the full tables.
Upload a mixed-language recording to the playground and see each word come back in the language it was spoken, with no code required.
Diarization across languages
Multilingual audio is usually multi-speaker audio. Turn on speaker_labels and the model separates who said what. AssemblyAI-STT-3.7 runs on a new speaker diarization model that is live for every language. On conversational recordings in nine non-English languages, cpWER is lower than on Universal-3.5 Pro in every one, from 23.5% to 22.1% on average. On English conversations it has the lowest diarization error rate of the pre-recorded APIs measured, 10.7% averaged over six datasets, ahead of ElevenLabs Scribe v3 (12.0%), Speechmatics Enhanced, Gladia Solaria-1, AWS Transcribe, Deepgram Nova-3 and Soniox. For how diarization works and how it's measured, see what is speaker diarization. For live audio, see streaming speaker diarization.
Sharpening accuracy with prompting and key terms
Two tools push multilingual accuracy further, and they work together:
promptis a short natural-language description of the audio, such as "Bilingual support call in English and Spanish about a billing dispute." Describe what is being transcribed; formatting instructions are ignored. To select languages, uselanguage_codes.keyterms_promptis a list of up to 1,000 words or phrases (six words per phrase at most) that you want spelled correctly: product names, people, jargon.
On AssemblyAI-STT-3.7, each is a $0.05/hr add-on. Key terms don't help equally in every language, so test them per language. We covered the failure modes in when keyterm prompting backfires.
Multilingual audio with PHI: Medical Mode and BAA
Healthcare audio is often multilingual, and it comes with protected health information (PHI).
Medical Mode. Add domain: "medical-v1" ($0.15/hr add-on) and the model gets medication names right, both generic and brand names, in English, Spanish, German and French. It works with all pre-recorded models, including AssemblyAI-STT-3.7, and it's also an add-on on both 3.7 realtime models. See the Medical Mode docs.
Compliance. AssemblyAI enables covered entities and their business associates subject to HIPAA to use the AssemblyAI services to process protected health information (PHI). AssemblyAI is considered a business associate under HIPAA, and we offer a standard Business Associate Addendum (BAA) that is required under HIPAA to ensure that AssemblyAI appropriately safeguards PHI. You can also redact PHI across audio and transcripts. See the medical solution page for details.
Where multilingual transcription goes next
Language count used to be the headline. The question that separates models now is what happens at the switch: the half-second where a speaker leaves one language and enters another, and the many more moments where an accented speaker stays in one language and the model has to resist flipping. With 32 languages code-switching natively in both recorded and live audio, you no longer route audio to a different model per caller, and you can design around the conversation. If you need translation on top, plan by mode. For live audio, the llm_gateway parameter on both 3.7 realtime models translates each final turn; real-time speech translation walks through it. For recorded audio, Speech Understanding Translation (86 target languages, $0.06/hr) runs on Universal-2, not on AssemblyAI-STT-3.7.
Multilingual transcription is available today through the Speech-to-Text API for pre-recorded audio and Streaming Speech-to-Text for real time. Pricing is on the pricing page.
Run AssemblyAI-STT-3.7 and AssemblyAI-STT-3.7-Realtime on your own multilingual audio. Start free, no credit card required.
Frequently Asked Questions (FAQs)
What is multilingual transcription and how does it work?
Multilingual transcription converts speech to text across more than one language, including recordings where speakers switch languages. The model detects the spoken language, transcribes each word in the language it was spoken, and keeps going when the language changes. AssemblyAI-STT-3.7 does this natively in 32 languages for recorded audio, and AssemblyAI-STT-3.7-Realtime does it in the same 32 languages live.
Is it multilingual or multi-language transcription?
Both terms describe the same thing: transcription that works across more than one language. "Multilingual" usually also implies the model can handle several languages in the same recording, including mid-sentence switches, which is what code switching means.
Which languages does AssemblyAI support in real time?
AssemblyAI-STT-3.7-Realtime and AssemblyAI-STT-3.7-Realtime-Agent both support 32 languages with native code switching, including English, Spanish, French, German, Hindi, Mandarin, Cantonese, Japanese, Korean, Arabic, Russian and Persian. No language parameter is required; language_codes steers the model when you know the languages.
Can AssemblyAI handle code switching?
Yes. AssemblyAI-STT-3.7 code-switches natively across 32 languages in recorded audio, with a 6.2% average word error rate on AssemblyAI's 16-dataset code-switching benchmark. AssemblyAI-STT-3.7-Realtime does the same in streaming, at 6.67% on five English language pairs. Neither needs extra configuration.
How do I transcribe languages AssemblyAI-STT-3.7 doesn't support?
Set speech_models to ["assemblyai-stt-3-7", "universal-2"]. AssemblyAI-STT-3.7 handles its 32 languages and Universal-2 covers the rest of the 99. With language_detection: true and only AssemblyAI-STT-3.7 in the list, there is no fallback.
Can I transcribe multilingual medical audio?
Yes. Add domain: "medical-v1" for medication names in English, Spanish, German and French on any pre-recorded model, including AssemblyAI-STT-3.7. AssemblyAI offers a BAA for customers processing PHI and can redact PHI in audio and transcripts.