AssemblyAI-STT-3.7: for pre-recorded audio, with 32 languages and powerful speaker diarization
Replacing the Universal family of models, AssemblyAI-STT-3.7 is available now on the pre-recorded Speech-to-Text API: 32 languages (up from 18), native code switching, and a new diarization model with 36% lower DER, all at the same $0.21/hr.
Today we're releasing AssemblyAI-STT-3.7, our new flagship model for pre-recorded audio.
It nearly doubles our language coverage, from 18 to 32. It handles code switching natively, so speakers can move from Urdu to English mid-sentence and still get an accurate transcript. It also ships with a new speaker diarization model that cuts speaker mix-ups substantially. Compared with Universal-3.5 Pro at launch, cpWER is down 20%, DER is down 36% and speaker-count errors are down 63%, and it runs 3–4× faster.
The price stays the same: $0.21/hr.
Try AssemblyAI-STT-3.7 Start free, or use your existing API key.
Leading accuracy across all 32 languages, 14 of them new
Most of the world's conversations don't happen in English. If your product transcribes meetings, calls or interviews, your users expect the same quality whether they speak Hindi, Portuguese or Finnish. A model that is strong in English and uneven everywhere else forces you to choose which customers get a good product. Language coverage only counts when the accuracy comes with it.
AssemblyAI-STT-3.7 adds 14 new languages: Urdu, Russian, Korean, Catalan, Galician, Romanian, Estonian, Persian, Cantonese, Afrikaans, Marathi, Zulu, Xhosa and Norwegian Nynorsk. Several of them are already among the most accurate languages on the model: Galician at 4.1% word error rate, Russian at 4.6% and Catalan at 5.1%. See the full list of supported languages in our docs.
The languages we already supported get better too. All 17 non-English languages supported on the previous model improve, and their mean error falls from 12.8% to 10.1%.
AssemblyAI-STT-3.7's mean error falls from 12.8% to 10.1% for already supported languages
Lower is better
Normalized word error rate; * character error rate for Japanese and Mandarin. 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. Languages are ordered by the AssemblyAI-STT-3.7 error rate.
Error rate on the 14 new languages
Lower is better
Normalized word error rate; * character error rate for Korean and Cantonese. The model was 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.
Code switching that follows the speaker, not the accent
Supporting 32 languages is only half the story, because people don't stick to one language per conversation. And the moment they switch can be the moment that matters most.
That's exactly where many speech-to-text systems slip. Some force the whole transcript into a single language, writing Catalan as Spanish or Galician as Portuguese. Others flip a few words or a whole sentence into another language even though the speaker never switched, especially when the speaker has an accent. The real challenge is doing both at once: following every genuine switch, and staying in one language when the speaker does.
AssemblyAI-STT-3.7 does exactly that. Earlier models would sometimes transcribe accented English in another language's script; AssemblyAI-STT-3.7 keeps it in English. Across our English benchmarks, stray characters from other scripts fell by 85%, from 161 to 24 per million characters.
That did not come at the cost of real code switching. On mixed-language audio, AssemblyAI-STT-3.7 is more accurate than Universal-3.5 Pro on six of seven language pairs, with the largest gains on English–Italian, English–German and English–Arabic, where errors fell by 16% to 20%. English–French is level. The model got better at following speakers who switch, and much better at leaving alone the ones who don't.
Code switching and language flips
Lower is better
Code-switching error rate (%)
AssemblyAI-STT-3.7, mixed-language audio, by language pair
Foreign-script characters
per million characters, English audio
Error rate is normalized word error rate; * character error rate for Mandarin and Japanese. Each pair is the mean over its code-switching datasets; the change column is relative to Universal-3.5 Pro on the same datasets. Foreign-script characters are non-Latin letters in the transcripts of English benchmark audio, per million output characters. Each panel has its own scale.
Who said it matters as much as what was said
When a sentence is credited to the wrong person, everything downstream can be wrong too, even if every word is right. That's why, alongside AssemblyAI-STT-3.7, we shipped a new speaker diarization model. It makes fewer attribution mistakes, even on short utterances, gets the number of speakers right far more often, and runs 3–4× faster. It's now live for every language on all of our speech models.
Most diarization systems decide who is speaking in each stretch of audio, then give every word the label of that stretch, even where the model wasn't sure. Our new diarization model takes the model's confidence into account when assigning each word to a speaker, so sentences stay intact while quick interjections like "yeah" and "mm-hmm" are still attributed to the right speaker.
Since our previous speaker diarization model update, all three measures have improved: cpWER fell from 29.3% to 23.5%, DER from 14.7% to 9.4%, and the average speaker-count difference from 0.57 to 0.21. The gain is particularly large on short clips: on 60-second clips, cpWER fell from 34.0% to 18.6%.
Lower diarization error with AssemblyAI-STT-3.7
Lower is better
cpWER
words plus speaker attribution
DER
diarization error rate
Speaker-count difference
predicted minus true speakers
Averages over English speaker-annotated conversation datasets. Full recordings use seven datasets; the 120-second and 60-second groups use the two of them that every system ran as clips, so the groups are not directly comparable with each other. Speaker-count difference is the predicted number of speakers minus the true number, averaged over files. Each panel has its own scale.
The improvement is not limited to English. On conversational recordings in nine other languages, cpWER is lower with AssemblyAI-STT-3.7 than with Universal-3.5 Pro in every one, falling from 23.5% to 22.1% on average.
Speaker-attributed word error rate (cpWER) by language
Lower is better
cpWER on conversational recordings with automatic language detection on both systems. * Scored on characters. Korean is new in AssemblyAI-STT-3.7, so it has no Universal-3.5 Pro figure and is not in the average.
The diarization step also runs 3–4× faster than before. If you already set
speaker_labels: true, you get it with no code change; it's live across all async speech
models.
More accurate word timestamps
Word timestamps are what captions, audio and video editors, and click-to-play search are built on. A boundary that is off by a tenth of a second is enough to clip a word or show a caption late. AssemblyAI-STT-3.7 uses a new word-timing method for English, so its word boundaries line up more closely with when each word is actually spoken.
We measured the distance between each predicted word boundary and a reference timing, on regular and on noisy English recordings. At word starts, the median error fell from 57 ms to 39 ms on regular audio and from 59 ms to 40 ms on noisy audio. Word end times also improved modestly. Timestamp improvements for other languages are coming soon.
Word timestamp error on English audio
Lower is better
Word start error (ms)
Word end error (ms)
Absolute difference between predicted and reference word boundaries, in milliseconds, over all matched words: median, 90th percentile (P90) and 95th percentile (P95).
You decide how the transcript is written
Different products need different transcripts. A meeting summary needs clean prose; a legal or research record needs every "um", repetition and false start; an audio or video editing tool needs a transcript that matches the recording word for word, so that cutting a word cuts the right sound. Universal-3.5 Pro already produces high-quality clean transcripts, but since its release, customers have told us they need precise control over these aspects of transcription.
AssemblyAI-STT-3.7 exposes these controls as two independent request parameters, so you choose the style of the output:
-
disfluencies:truekeeps fillers, repetitions and false starts as spoken, for verbatim transcripts (the default);falseremoves them for a clean transcript. -
format_text:truewrites numbers, currencies, dates and punctuation in written form (the default);falsereturns spoken-form text. For example:- Times and amounts: "twelve fifty" rather than "12:50" or "$12.50".
- Email addresses: "jane dot doe at example dot com" rather than "jane.doe@example.com".
- URLs: "example dot com slash pricing" rather than "example.com/pricing".
- Phone numbers: "four one five, five five five, zero one three two" rather than "415-555-0132".
On AssemblyAI-STT-3.7 both controls default to true, giving verbatim, written-form
transcripts. This is a change from Universal-3.5 Pro, where disfluencies were off by default. For clean
transcripts without fillers, set disfluencies: false.
We score verbatim fidelity with four F1 metrics, each comparing the events in the transcript with those in a verbatim reference after aligning the two word by word. With disfluencies on, the model keeps most fillers and repetitions; with disfluencies off it removes them, which is what the low scores for that setting mean: the setting working as designed, not a weaker model.
Verbatim fidelity, F1 (%)
higher is more verbatim, lower is cleaner
Mean over the benchmark datasets whose references label each event reliably. Disfluency F1 covers fillers, hesitations, backchannels and restart fragments; contraction F1 covers colloquial reduced forms such as "gonna" and "'em"; repetition F1 covers immediately repeated words; stutter F1 covers word fragments marked in the reference. Each F1 combines recall, the share of reference events the transcript keeps, and precision, the share of transcript events that are in the reference, as their harmonic mean.
Measured against the field
Beating our own previous model is only half of what you need to know. When you choose a speech-to-text provider, the question is how it compares with the alternatives. So we compared AssemblyAI-STT-3.7 with the pre-recorded models of other providers on the same recordings, each through its production API with default settings, and scored every transcript the same way, using both publicly accessible benchmarks and our own internal ones.
Public benchmarks
We start with benchmarks that anyone can inspect and reproduce: the datasets are public, and so is the scoring.
Open ASR Leaderboard
The Open ASR Leaderboard is a public benchmark with an open evaluation harness. We ran AssemblyAI-STT-3.7 through that harness and placed the result among the leaderboard's published entries.
AssemblyAI-STT-3.7 would rank 5th of 70 entries on English short-form, 3rd of 42 on multilingual and 2nd of 15 on long-form. That is a step up from Universal-3.5 Pro on every board, which ranks 16th, 4th and 6th. No single model is ahead of AssemblyAI-STT-3.7 on all three boards.
English short-form 5th of 70 mean WER over seven public sets
Ahead of it: zoom/scribe_v2.1_pro 4.19 · OtterAI/speech-v3.1 4.57 · yoob/yoob-asr-pro-v1 4.77 · microsoft/azure-speech-09-2026 4.87
Multilingual 3rd of 42 mean WER over six languages
Ahead of it: microsoft/azure-speech-09-2026 2.53 · elevenlabs/scribe_v2 2.76
Long-form 2nd of 15 mean WER over the long-form sets
Ahead of it: elevenlabs/scribe_v2 8.45
AppTek call-center dialogues (Multi-Accent Long-Form Benchmark)
AppTek's call-center dialogues are a public dataset of English customer-service calls recorded across 14 accents, published with its own scoring script. We ran every system with automatic language detection and scored the transcripts with that script. With default settings, AssemblyAI-STT-3.7 has a 7.44% word error rate, close to ElevenLabs Scribe v3 at 7.34% and down from 8.35% for Universal-3.5 Pro. The dataset's references spell out every number, so we also ran the evaluation with text formatting turned off, which returns spoken-form text. With that setting, the error rate falls to 7.05%.
Internal benchmarks
The public benchmarks above measure transcription accuracy. They do not cover speaker diarization or code switching, so for those we use our own benchmarks. AssemblyAI-STT-3.7 has the lowest error rate on both. AssemblyAI-STT-3.7 and ElevenLabs Scribe v3 were measured in November 2026; the other providers' results come from the September and October 2026 runs behind our benchmarks page.
Diarization
On speaker diarization for English conversations, AssemblyAI-STT-3.7 has the lowest diarization error rate, 10.7%, ahead of ElevenLabs Scribe v3 at 12.0%. The other systems shown are at 15% or higher.
Code switching
Code switching is where the field spreads out most. AssemblyAI-STT-3.7 has the lowest average error rate, 6.2%, on audio that mixes English with German, Spanish, French, Italian or Portuguese, and the lowest error rate on 12 of the 16 datasets. ElevenLabs Scribe v3 follows at 6.8% and leads on the other four; Soniox is at 10.2%. The remaining systems are at 17% or higher, and several are above 40%.
Pricing
AssemblyAI-STT-3.7 is available for $0.21/hr of audio, unchanged from the previous model. Your current rate for pre-recorded API access stays the same. See our pricing.
Get started
AssemblyAI-STT-3.7 is available now with your existing API key. Change the model name, run your own recordings through it, and start seeing improvements today.
Try it on your own audio Use your existing API key.
Frequently asked questions
How do I switch from Universal-3.5 Pro?
Set speech_models to ["assemblyai-stt-3-7"]. There is no new endpoint. For
the full steps, see the
AssemblyAI-STT-3.7 migration guide.
import requests
response = requests.post(
"https://api.assemblyai.com/v2/transcript",
headers={"authorization": "YOUR_API_KEY"},
json={
"audio_url": "https://example.com/meeting.mp3",
"speech_models": ["assemblyai-stt-3-7"],
"speaker_labels": True,
},
)
print(response.json()["id"]) If I don't set speech_models, will I get the latest model?
It depends on your account:
- New accounts created on or after November 17, 2026: requests that don't set
speech_modelsalready useassemblyai-stt-3-7. - Existing accounts that default to
universal-3-5-prooruniversal-3-pro: starting December 8, 2026, requests that don't setspeech_modelswill useassemblyai-stt-3-7automatically. This changes the default model only;universal-3-5-proanduniversal-3-proare not being retired on that date. - Existing accounts that default to
universal-2: these are not upgraded automatically.
If you do set speech_models, your requests keep using the model you specified until that
model is retired. To test AssemblyAI-STT-3.7 and switch to it, see the
AssemblyAI-STT-3.7 migration guide.
What about languages outside the 32?
Set speech_models: ["assemblyai-stt-3-7", "universal-2"] for support across 99 languages. If
you set language_code to a language AssemblyAI-STT-3.7 doesn't support, the request goes to
Universal-2. The speech_model_used field tells you which model ran. See
transcribe all 99 languages by combining speech models
in our docs.
Is Universal-3.5 Pro still available?
Universal-3 Pro and Universal-3.5 Pro are deprecated, with a retirement date to be announced. Once they
are retired, requests specifying universal-3-pro or universal-3-5-pro will route
to AssemblyAI-STT-3.7 automatically. Universal-2 remains available. See
model versions and deprecations
in our docs.