Introducing AssemblyAI-STT-3.7: next generation transcription models for live conversations, voice agents, and recordings
Introducing the new frontier in transcription models: AssemblyAI-STT-3.7-Realtime for live conversation, AssemblyAI-STT-3.7-Realtime-Agent for conversations with voice agents, and AssemblyAI-STT-3.7 for recorded audio. Our new models provide cutting-edge accuracy and latency on real voice workloads.
Today we're releasing AssemblyAI-STT-3.7, our next generation of transcription models: AssemblyAI-STT-3.7-Realtime for live conversations, AssemblyAI-STT-3.7-Realtime-Agent for conversations with voice agents, and AssemblyAI-STT-3.7 for recorded audio. All three hold up on the hardest audio, at very low latency.
Key details
| Detail | AssemblyAI-STT-3.7-Realtime | AssemblyAI-STT-3.7-Realtime-Agent | AssemblyAI-STT-3.7 |
|---|---|---|---|
| Built for | Live conversations: meetings, calls, captions | Voice agents | Recorded audio |
| API | Realtime Speech-to-Text API | Realtime Speech-to-Text API | Pre-recorded Speech-to-Text API |
| API parameter | speech_model= | speech_model= | speech_models= |
| Languages | 32 in one stream, with automatic detection | 32 in one stream, with automatic detection | 32, with automatic detection and an optional 99-language fallback |
| Speaker labels | Similar diarization accuracy as pre-recorded, but labels are applied the moment a turn ends | Focuses on the person talking to the agent. Background speech cut by about half. Time to final transcript in 86 ms. | Lowest diarization error rate of the pre-recorded APIs we measured |
| Price | $0.30/hr | $0.45/hr | $0.21/hr |
| Replaces | Universal-Streaming and Universal-Streaming Multilingual | Universal-3.6 Pro Realtime | Universal-3.5 Pro (pre-recorded) |
AssemblyAI-STT-3.7-Realtime
Customers expect a transcript on screen while the meeting happens and speaker-attributed notes before the call ends. Many notetakers still batch audio to the pre-recorded API anyway, because live speaker labels used to settle only after the call and streaming cost more than the recorded pipeline. AssemblyAI-STT-3.7-Realtime changes both.
AssemblyAI-STT-3.7-Realtime is built for live conversation: meetings, calls, and captions. From
the beginning of a new session, speaker labels are final the moment a turn ends and are nearly
as accurate as labels from pre-recorded audio. There is no revision pass and no
SpeakerRevision message to wait for. Partials arrive about once a second, and turns
close when there is enough to show, which is the behavior a live view wants.
Measured against the field on the same recordings, it has the lowest live diarization error rate of every realtime provider we measured on telephone calls and on multi-party meetings recorded on a single distant microphone, ahead of ElevenLabs Scribe v3 Live and Meta Muse Voice Transcribe.
Live speaker labels on telephone calls
Lower is better
Live speaker labels on multi-party meetings
Lower is better
On short-form English and on code-switched audio, AssemblyAI-STT-3.7-Realtime leads every other provider we measured. Short-form English word error rate is 3.23%, ahead of ElevenLabs Scribe v3 Live, AWS Transcribe Streaming and Deepgram Flux. On code-switched audio it is 6.67%, first on all five language pairs.
Short-form English word error rate against the field
Lower is better
Word error rate on code-switched audio against the field
Lower is better
Three recordings from the code-switching set, a sentence in one language followed by a sentence in English.
Spanish, then English. Neither other provider gets the English sentence after the switch; ours writes both sentences.
Durante la temporada, Liriano estableció marcas personales en victorias y ponches. It infuriates me, and I don't know why.
Durante la temporada. Liriano estableció marcas personales en victorias y ponches. It infuriates me, and I don't know why.
Durante la temporada, Liriano estableció marcas personales en victorias y ponches. Y 2020 en la de .
Durante la temporada, Lidiano estableció marcas personales en victorias y ponches. Ich versuche jetzt nicht, meine Kunden online
It serves all 32 languages in one stream with automatic language detection, which merges Universal-Streaming and Universal-Streaming Multilingual into one model. If you kept a language router in front of two streaming integrations, you can retire it. It is the most accurate realtime API we measured on Spanish, French, German and Galician, and it stays in the caller's language: words emitted in English on non-English audio fall to 0.033%, 44% below Soniox Realtime v6, a tenth of Google Gemini 3.5 Transcribe Live. Across the 18 Common Voice languages, both AssemblyAI-STT-3.7 realtime models are more accurate than any other realtime API that covers all 18, ahead of ElevenLabs Scribe v3 Live, Soniox Realtime v6 and Google Gemini 3.5 Live; no Deepgram model covers all 18. Key terms and a 400-character prompt steer product names and vocabulary the same way they do on the pre-recorded model.
What it doesn't do: turn-taking. If anything on your side waits on the model to decide the caller is done, use AssemblyAI-STT-3.7-Realtime-Agent.
Priced at $0.30 per hour, one-third less than the agent model, so the per-hour cost stops being the reason to batch your live conversation transcription workloads.
AssemblyAI-STT-3.7-Realtime-Agent
A caller says "no" and the agent doesn't respond. A caller asks for an appointment time and the agent books the wrong one. At a drive-thru, the kids in the back seat shout an order and the agent takes it. Voice agent builders told us these three problems break more calls than anything else. AssemblyAI-STT-3.7-Realtime-Agent is built to solve them, at low latency.
Latency: the final transcript in 86 ms
In a cascaded agent, speech-to-text latency is the time from the caller's last word to the final transcript, plus the time to decide the caller is done. AssemblyAI-STT-3.7-Realtime-Agent returns the final transcript 86 ms after speech ends. Only 1.0% of words change between partial and final, down from 3.2%, so what the agent acts on early is what it gets at the end. With Gemma 4 31B through AssemblyAI's LLM Gateway and AssemblyAI-TTS-1, end-to-end turn latency is regularly around 250 ms.
Speed against accuracy on voice-agent calls
Lower is better on both axes
Numbers are where agents fail loudest, and where this model moved most. On 12,460 English voice-agent calls, number errors on Voice Code Bench fall 80% against Universal-3.6 Pro Realtime, so the amount, the account number, and the confirmation code come through. Short replies, the one- and two-word answers that decide a call, are wrong 29% less often and dropped 38% less often. Against the field, entity error on those short utterances is 12.0%, 28% below the next provider, and ahead of Soniox Realtime v6, ElevenLabs Scribe v3 Live and Deepgram Nova-3.
Here are two calls from the voice-agent set: an order code and a name followed by two short replies, and a support contact with an account number and an e-mail address. Words that differ from the reference are underlined.
A UK-accented caller in heavy noise. One other provider loses “Uber Eats”, the other changes “Jane Davis”; ours keeps both.
Please check my Uber Eats order, ORD-652895, for Jane Davis. Not yet. No, please issue a refund instead.
Please check my Uber Eats order, ORD 652895, for Jane Davis. Not yet. No, please issue a refund instead.
Please check my delivery order, ORD652895, for Jane Davis. No. Please issue a refund instead.
Please check my Uber Eats order. ORD652895, for Jane David. Not yet. No. Please issue a refund in
Background speech breaks agents outright. The model focuses on the person talking to the agent: words added while the caller was silent fall 47%, so the order comes from the driver, not the back seat. In heavy background noise its word error rate is 5.32%, ahead of Soniox Realtime v6, ElevenLabs Scribe v3 Live and Deepgram Nova-3. Background speech suppression works without configuration for most call environments, but for cases where a second talker is close to the microphone we recommend using the near-field voice focus feature.
Two calls in heavy noise. On one, both other providers write a background talker's words as if the caller said them; on the other, both change the e-mail address the caller reads out.
Heavy background noise. Both other providers change the e-mail address the caller reads out; ours gets every word.
Please message Anna from Amazon IT and ask if Slack will be fixed soon. No, it's anna.tech@amazon.com.
Please message Anna from Amazon IT and ask if Slack will be fixed soon. No, it's anna.tech@amazon.com.
Hey, this is Adam from Amazon IT, and I asked you flat will be 16. No, it's adam.ick@amazon.com.
Message Anna from Amazon IT and ask you flat will be 16. No. It's Anna.pick@amazon.com.
Fewer errors on voice-agent calls
Reduction vs Universal-3.6 Pro Realtime
Callers who speak English as a second language are where every provider makes the most entity errors. AssemblyAI-STT-3.7-Realtime-Agent has the lowest rate in every one of the 12 speaker groups we measure. On Indian-accented callers it is 17.6%, ahead of Soniox Realtime v6, ElevenLabs Scribe v3 Live and Deepgram Nova-3, and its 5.19% word error rate on Indian-accented English is the lowest of 30 realtime models.
Two calls from Indian-accented callers.
An Indian-accented caller under heavy noise. Ours writes the e-mail address exactly; both other providers misspell the name in it.
Can you check if emails to johnjones@gmail.com are still delayed? Just John. No, I said he's on Amazon WorkMail.
Can you check if emails to johnjones@gmail.com are still delayed? Just John. No, I said he's on Amazon WorkMail.
Can you check if emails to johnjohns@gmail.com are still delayed? Yes, John. No, I said he's on Amazon work mail.
Can you check if emails to JohnJonas@gmail.com are still delayed Just John. No. I said he's on Amazon Work Meal.
Against the other providers we measured on the same 12,460 calls, with no key terms or context for anyone, AssemblyAI-STT-3.7-Realtime-Agent has the lowest word error rate, the lowest entity error rate, and the fewest dropped short replies, ahead of Soniox Realtime v6, ElevenLabs Scribe v3 Live and Deepgram Nova-3. No other provider has both less background speech and fewer dropped short replies.
Word error rate on voice-agent calls
Lower is better
Entity error rate on voice-agent calls
Lower is better
Two independent benchmarks ran AssemblyAI-STT-3.7-Realtime-Agent on their own audio. Voice Arena scores every model on a private conversational set that no vendor can train against, and ranks AssemblyAI-STT-3.7-Realtime-Agent the #1 streaming model for US English at 2.45% semantic WER, ahead of Fish Transcribe Stream 1 (2.75%) and ElevenLabs Scribe v3 Live (2.89%), with Indian English also ahead of Gemini 3.5 Transcribe Live. Speko scores streaming models on word error rate, end-of-turn latency, and time to first token at real-time pace, and AssemblyAI-STT-3.7-Realtime-Agent comes in ahead of ElevenLabs Scribe v3 Live, Deepgram Flux, Soniox Realtime v6, Fish Transcribe Stream 1, and Cartesia Ink-2 on all three.
The price is $0.45/hr, the same as Universal-3.6 Pro Realtime.
Read the full voice-agent benchmark and methodology in the AssemblyAI-STT-3.7-Realtime-Agent research post.
AssemblyAI-STT-3.7
An accurate transcript is the foundation for everything built on it: meeting notes, call coaching, clinical records, search, the inputs to an LLM feature. Small word errors wash out in post-processing. The wrong speaker on a line, a sentence in the wrong script, an amount or a name transcribed wrong can't be undone downstream. AssemblyAI-STT-3.7 gets these right. It knows who spoke, stays in the language they spoke, and captures the entities that matter.
Measured against the field on the same recordings, AssemblyAI-STT-3.7 has the lowest diarization error rate of any pre-recorded API we measured, ahead of ElevenLabs, with no other provider below 15%. On audio that switches between English and another language it has the lowest average error rate, and the lowest on 12 of the 16 datasets we test.
Diarization error rate on recorded meetings and calls
Lower is better
AssemblyAI-STT-3.7 runs on a new speaker diarization model, which is also rolling out across all of our models and every supported language. It makes fewer mistakes, gets the number of speakers right far more often, and runs faster. The biggest gains are on short clips, which give the model the least audio to learn each voice from, so speaker labels now hold up on voice notes and short call snippets.
Your customers speak more languages as you grow, and many switch languages mid-conversation or mid-sentence. A model that handles that code-switching usually pays for it with language flips in conversations that never switched, like a stray line of Devanagari in accented English. AssemblyAI-STT-3.7 avoids the tradeoff: it leads the field on code-switched audio and cuts the language flip rate by 80 to 90% against our previous models, so every word comes back in the language and script it was spoken in.
Word error rate on code-switched audio
Lower is better
Three recordings from the code-switching sets, an English sentence and a second language in one file. Words that differ from the reference are underlined.
An English sentence, then a German one. ElevenLabs and Deepgram both return only the German sentence; AssemblyAI-STT-3.7 writes both.
And so, despite the fact that he was a Roman centurion, he converted to their faith. Dies ist meine erste Frage.
And so, despite the fact that he was a Roman centurion, he converted to their faith. Dies ist meine erste Frage.
Dies ist meine erste Frage.
Dies ist meine erste Frage.
Language coverage nearly doubles, from 18 to 32: Urdu, Russian, Korean, Catalan, Galician, Romanian, Estonian, Persian, Cantonese, Afrikaans, Marathi, Zulu, Xhosa and Norwegian Nynorsk are new, and Galician (4.1% word error), Russian (4.6%) and Catalan (5.1%) are already among the model's most accurate. The 17 non-English languages the previous model covered improve from 12.8% to 10.1% mean error.
How the transcript is written is up to you. A meeting summary wants clean prose; an editor needs
the recording word for word. Two parameters control this, and both improve:
disfluencies keeps or removes fillers and false starts, and
format_text writes numbers and dates in written form ("$1,250", "March 3") or
spoken form. Both now default to true, so out of the box you get fillers kept and
numbers written out. Universal-3.5 Pro used different defaults, so set both explicitly when you
switch.
There is no automatic fallback to Universal-2 unless you list it, and transcripts are verbatim unless
you turn disfluencies off. For languages outside the 32, add universal-2 to
speech_models and you're covered for 99.
Read the full technical write-up in the AssemblyAI-STT-3.7 research post.
Handle, whose users want a meeting summary they can read or share the moment the call ends, builds on AssemblyAI-STT-3.7:
We've been building around voice for the last two years, and it's a big part of where we're headed. Our users want to finish a meeting and have a summary they can read or share almost instantly, and they rely on us to capture exactly what was said. That's a high bar for both speed and accuracy, and neither matters without stability and reliability. We found all of it with AssemblyAI-STT-3.7. The team made it easy, and they share our beliefs about the future of voice. We're excited about what's next for Handle, where voice will play a major role.
Which model to use
| If your audio is | Use | Why |
|---|---|---|
| A live stream that ends in a transcript | AssemblyAI-STT-3.7-Realtime | Speaker labels final at turn end, 32 languages in one stream |
| A live stream your voice agent answers | AssemblyAI-STT-3.7-Realtime-Agent | Short replies, numbers, and partials your voice agent can act on |
| A file, a recording, anything you can wait for | AssemblyAI-STT-3.7 | Highest accuracy on recorded speech across 32 languages at the lowest cost |
A person waiting on a reply means the agent model. A transcript nobody has to answer means the realtime model.
Pricing
AssemblyAI-STT-3.7-Realtime is $0.30/hr. AssemblyAI-STT-3.7-Realtime-Agent is $0.45/hr, unchanged from Universal-3.6 Pro Realtime. AssemblyAI-STT-3.7 is $0.21/hr, unchanged from Universal-3.5 Pro, and your current pre-recorded rate carries over. See the pricing page.
Get started
The samples below use the Python SDK, as in the streaming and pre-recorded quickstarts.
Realtime, for conversation
import os
import time
from assemblyai.streaming.v3 import (
RealTimeEvents,
RealTimeParameters,
RealTimeTranscriber,
RealTimeTranscriberOptions,
TurnEvent,
)
def on_turn(client: RealTimeTranscriber, event: TurnEvent):
if event.end_of_turn:
print(f"Speaker {event.speaker_label}: {event.transcript}")
client = RealTimeTranscriber(
RealTimeTranscriberOptions(terminate_timeout=30.0),
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",
speaker_labels=True,
)
)
with open("meeting.pcm", "rb") as f: # 16 kHz, 16-bit mono PCM
while chunk := f.read(3200): # 100 ms chunks, sent at live speed
client.stream(chunk)
time.sleep(0.1)
client.disconnect(terminate=True)Realtime, for conversations with agents
# Same code as above. Change the model and drop speaker_labels:
client.connect(
RealTimeParameters(
sample_rate=16000,
speech_model="assemblyai-stt-3-7-realtime-agent",
)
)Pre-recorded
import os
from assemblyai.prerecorded.v2 import Transcriber, TranscriptionConfig
transcriber = Transcriber(api_key=os.environ["ASSEMBLYAI_API_KEY"])
config = TranscriptionConfig(speech_models=["assemblyai-stt-3-7"], speaker_labels=True)
transcript = transcriber.transcribe("https://assembly.ai/wildfires.mp3", config=config)
for utterance in transcript.utterances:
print(f"Speaker {utterance.speaker}: {utterance.text}")The docs carry the migration guide and the model names table.
Frequently asked questions
Why three models, and why the new names?
Recorded audio, live conversation, and conversation with an agent reward different things, and a model tuned for one gives up something on the others. The name makes the choice explicit: the surface, then the use case. The Universal models keep running; the model names table maps every old name to the one to use now.
How do I switch from Universal-3.5 Pro for pre-recorded audio?
Set speech_models to ["assemblyai-stt-3-7"]. There is no new endpoint. Two
defaults changed: no automatic fallback to Universal-2 unless you list it, and verbatim transcripts by
default. See the migration guide.
How do I switch from Universal-Streaming or Universal-3.6 Pro Realtime?
Set speech_model to assemblyai-stt-3-7-realtime for live conversations or
assemblyai-stt-3-7-realtime-agent for voice agents. The endpoint and message protocol are
unchanged. On the
realtime model, Universal-Streaming's turn-detection parameters are accepted and ignored, and speaker
labels no longer send revisions.
What about languages outside the 32?
For pre-recorded audio, set speech_models: ["assemblyai-stt-3-7", "universal-2"] for 99
languages in total. If you pin language_code to a language AssemblyAI-STT-3.7 doesn't
support, the request routes to Universal-2, and speech_model_used tells you which model ran.
Are the Universal models being retired?
Universal-3 Pro and Universal-3.5 Pro for pre-recorded audio are deprecated, with a retirement date to be announced; after retirement, requests naming them route to AssemblyAI-STT-3.7 automatically. Universal-3.5 Pro Realtime will be retired on February 1, 2027. Universal-2 remains available.
Where can I read how this was measured?
Pre-recorded figures compare AssemblyAI-STT-3.7 with Universal-3.5 Pro on our internal evaluation sets at default settings. Realtime figures use five meeting and call sets (466 recordings, 83 hours); voice-agent figures use 12,460 English role-play calls with paired 95% intervals. Competitor figures come from the same recordings through the same harness, each vendor's realtime API at its defaults, with no key terms or context for anyone. The research posts carry the full tables and the benchmarks page the methodology.