Three models, three prices, 32 languages in one stream: AssemblyAI replaced its whole transcription line, and three new defaults will bite you on upgrade
The short version
On November 18, AssemblyAI retired the idea that one transcription model can cover everything. It split its line into three: AssemblyAI-STT-3.7 for recorded audio at $0.21 per hour, -Realtime for live captions at $0.30, and -Realtime-Agent for software that has to answer the caller at $0.45. All three cover 32 languages with automatic detection. Prices did not move. The model names, the defaults, and the fallback behavior all did. The hard part of this release is not how good the models are — it is that three silent default changes will quietly rewrite what your pipeline outputs.
What this actually is
Keep transcription and voice assistants separate. All three models do one thing: turn sound into text. They do not answer questions and they do not make decisions. They emit transcripts with speaker labels. What differs is where the audio comes from and how long you can afford to wait.
- Recorded (
AssemblyAI-STT-3.7): files, meeting recordings, podcasts. Waiting tens of seconds is fine. - Realtime (
-Realtime): live meeting captions and call transcripts, where words need to appear as they are spoken. - Agent (
-Realtime-Agent): your voice assistant needs that text immediately to decide what to do next, so this model also has to judge whether the caller has finished speaking.
In the past these three jobs were stitched together from different Universal models, with a naming scheme messy enough to require a mapping table. This time AssemblyAI made the rule explicit: surface first, then use case. The API endpoints and the message protocol are unchanged. Only the value you put in speech_model or speech_models changes.
Why it matters right now
Because this is not one more SKU. A whole generation is being replaced while the previous generation enters its retirement countdown. The official changelog spells it out:
- On the recorded side, Universal-3 Pro and Universal-3.5 Pro are deprecated with a retirement date still to be announced. Once they retire, requests naming them route to STT-3.7 automatically.
- On the streaming side, Universal-3.5 Pro Realtime retires on February 1, 2027.
- Universal-2 stays, as the fallback for anything outside the 32 languages.
So if your code has an old model name hard-coded, you have until next February at the latest. Promptwire, checking the lifecycle documentation, added a nearer date: from December 8, 2026, existing accounts that omit speech_models and currently default to Universal-3.5 Pro or Universal-3 Pro will be moved to STT-3.7. Accounts that default to Universal-2 are excluded.
This is not a question of whether to upgrade. It is a timetable for when you have to deal with it.
Realtime: speaker labels settle instantly, but it is not the fastest
Live transcription has carried an awkward compromise: words stream out, but "who said this" often only firms up after the call ends. That is why plenty of notetakers still batch audio into the recorded API. STT-3.7-Realtime attacks exactly that. AssemblyAI says labels are final the moment a turn closes, nearly as accurate as recorded-audio diarization, with no revision pass and no SpeakerRevision message to wait for.
The official blog post's own head-to-head numbers, lower being better:
- Live diarization error on telephone calls: STT-3.7-Realtime 10.46%, ElevenLabs Scribe v3 Live 10.78%, Speechmatics 14.58%, Meta Muse 16.44%, Soniox Realtime v6 18.35%, Azure 20.21%, AWS 27.89%, Deepgram Nova-3 30.46%.
- Multi-party meetings, one distant microphone, up to seven speakers: 6.30%, Meta Muse 7.34%, Scribe v3 Live 9.63%, Azure 19.19%, AWS 23.79%.
- Short-form English word error rate: 3.23%, AWS 3.93%, Scribe v3 Live 4.18%, Gemini 3.5 Transcribe Live 5.04%, Nova-3 7.46%.
- Code-switched audio, English paired with Spanish, German, French, Portuguese and Italian: 6.67%, Scribe v3 Live 8.32%, AWS 9.13%.
Worth noting: the post says plainly that this model does not do turn-taking. If your stack waits for the model to declare the caller finished, you need the agent model. That limitation is stated in the body text, not buried.
Realtime-Agent: 86 milliseconds to final, at the cost of being slower than the fastest
This is the most practical of the three, because it targets the three ways voice agents actually fail: the caller says "no" and the agent keeps talking, the caller gives a time and the agent books the wrong slot, kids in the back seat at a drive-thru shout an order and the agent takes it.
- Latency: the body text says the final transcript lands 86 ms after speech ends. The scatter plot on the same page labels the same model at 88 ms with 4.53% word error rate. Two milliseconds apart, no effect on the conclusion, but it tells you a human typed these numbers.
- Stability: only 1.0% of words change between partial and final, down from 3.2%. That is what decides whether the sentence your agent already acted on survives to the end.
- On 12,460 English voice-agent calls, versus the previous Universal-3.6 Pro Realtime: number errors drop from 11.28% to 2.27% (-80%), entity misses from 23.28% to 11.00% (-53%), words inserted while the caller is silent from 2.53 to 1.34 per thousand words (-47%), dropped short replies from 0.29% to 0.18% (-38%), wrong short replies from 1.35% to 0.96% (-29%), false end-of-turn events down 12%.
- Word error rate against the field: 4.53%, Soniox Realtime v6 6.02%, Scribe v3 Live 7.38%, Meta Muse 7.48%, Gemini 3.5 Live 8.48%, Nova-3 8.64%.
- Heavy background noise: 5.32%. Indian-accented English: 5.19%, which AssemblyAI calls the lowest of 30 realtime models.
- End to end: with Gemma 4 31B through AssemblyAI's LLM Gateway plus AssemblyAI-TTS-1, a conversational turn lands around 250 ms.
Read the trade-off carefully. On that same speed-versus-accuracy chart, four models reach the final faster: Soniox Realtime v6 at 56 ms, Cartesia Ink Whisper at 68 ms, Deepgram Nova-3 at 69 ms, Meta Muse at 76 ms. AssemblyAI's framing is that every provider reaching the final sooner makes at least 30% more word errors. That is a comparison condition the vendor set, not an industry consensus, but at least it puts speed and accuracy on one chart and lets you choose. Cartesia's dot is 68 ms paired with 15.17% word error rate — unusable for reading out an order number.
Two outside checks exist, and both deserve a discount. Voice Arena ranks Realtime-Agent first among streaming models for US English at 2.45% semantic word error rate, ahead of Fish Transcribe Stream 1 at 2.75% and Scribe v3 Live at 2.89%. Speko ranks it ahead of Scribe v3 Live, Deepgram Flux, Soniox Realtime v6, Fish Transcribe Stream 1 and Cartesia Ink-2 on word error rate, end-of-turn latency and time to first token. How those sets were sampled, and how large they are, is not explained in this post.
Recorded: 18 languages become 32, diarization rebuilt
The recorded model is the least glamorous and the one most teams actually run: meeting notes, interview cleanup, subtitles.
- Diarization: a new model across six English datasets averages 10.7% diarization error rate. Scribe v3 is 12.0%, xAI Grok STT 15.4%, Voxtral Mini Transcribe 2 15.9%, Speechmatics 16.2%, AWS 16.9%, Nova-3 17.4%, Soniox stt-async-v5 24.2%. The company says the biggest gains land on short clips — voice notes and 30-second call fragments, the material that gives a model the least audio to learn a voice from.
- Languages go from 18 to 32: Urdu, Russian, Korean, Catalan, Galician, Romanian, Estonian, Persian, Cantonese, Afrikaans, Marathi, Zulu, Xhosa and Norwegian Nynorsk are new. The 17 non-English languages the previous model covered improve from 12.8% to 10.1% mean error. Among the additions, Galician at 4.1%, Russian at 4.6% and Catalan at 5.1% already rank among its strongest.
- Code-switching: 6.2% word error rate, against Scribe v3 at 6.8%, Soniox at 10.2%, Azure at 17.4%, AWS at 22.9%, GPT-4o Transcribe at 44.7%, Nova-3 at 49.4% and Gladia at 50.4%. At the same time, language flips — a stray line of Devanagari inside accented English — fall 80% to 90%. Those two properties used to fight each other: loosen language detection to handle switching and you get phantom script in monolingual audio.
The money: three prices, two different billing units
- Recorded: $0.21 per audio hour, unchanged from Universal-3.5 Pro.
- Realtime: $0.30 per hour, which AssemblyAI pitches as a third cheaper than the agent model — the point being that per-hour cost stops being your reason to batch live audio.
- Realtime-Agent: $0.45 per session hour, unchanged from Universal-3.6 Pro Realtime.
One detail is easy to miss, and Promptwire flagged it while checking the pricing page: recorded audio is billed by audio duration, streaming is billed for the time the WebSocket session stays open. A connection sitting there with nobody talking is still burning money. Cost estimates cannot be built by multiplying the minutes of human speech by the unit price.
Pitfalls: three defaults will change your output
This is the part worth pinning. None of these throw an error. They just change what comes out.
- No automatic fallback to Universal-2 anymore. Previously the system quietly caught languages outside coverage. Now, with
speech_models: ["assemblyai-stt-3-7"]alone, anything outside the 32 simply is not supported. You have to listuniversal-2explicitly to reach 99. AssemblyAI exposes aspeech_model_usedfield telling you which model actually ran — sample it after rollout. disfluenciesandformat_textnow default to true. Out of the box you get fillers and false starts preserved, and numbers and dates written in written form ("$1,250", "March 3"). Subtitle teams that want clean prose, and editors who need a verbatim record, must turn these off explicitly. The old defaults were different, so this is a behavior change, not a no-op.- The endpoint is the same, the semantics are not. The realtime model accepts Universal-Streaming's turn-detection parameters and then ignores them, and speaker labels no longer send revisions. If downstream code is waiting for
SpeakerRevision, it will wait forever.
Good fit, and not a good fit
A good fit if you:
- Build meeting notes or notetakers and want live captions and speaker labels together, without batching audio first.
- Run voice agents or contact centers and are tired of wrong digits, order codes and email addresses. This is where the gains are most concrete.
- Operate multilingually, especially English mixed with another language on support calls.
Not a good fit if you:
- Need the model to decide who speaks next. Realtime explicitly does not do turn-taking; you need the agent model at 50% higher price.
- Want open weights to self-host. All three are API-only.
- Work outside the 32 languages. You can, but only by listing Universal-2 explicitly and accepting the previous generation.
- Are choosing purely on outside leaderboards. Apart from the Voice Arena and Speko mentions, every head-to-head number here came from AssemblyAI running the same harness. Promptwire states outright that it did not independently benchmark the models.
How to use it: five steps before and after the upgrade
Treat this as the action list. You do not need to change every system at once; start with the first step and work down.
- First, find out which model your requests actually use. Not what your code says — what
speech_model_usedreturns. - If you omit
speech_modelsand currently default to Universal-3.5 Pro, put December 8 on the calendar. - Run three kinds of material through it: a clean single-speaker interview, a noisy multi-party meeting, and a recording that changes language mid-file. Compare word errors, speaker attribution, and how much human cleanup each one actually costs you.
- Pin
disfluenciesandformat_textexplicitly. Do not rely on defaults. - For streaming, include idle connection time in your cost model.
Sources
- AssemblyAI launch post (November 18, 2026): https://www.assemblyai.com/blog/introducing-assemblyai-stt-3-7
- AssemblyAI changelog, entry dated November 18, 2026: https://www.assemblyai.com/changelog
- Promptwire third-party analysis (November 19, 2026): https://www.promptwire.org/articles/assemblyai-stt-3-7-models-pricing-transcription-migration