AssemblyAI-STT-3.7 ships as three models, and they split on two questions: is the audio already recorded, and if it's live, does a person read the transcript or does software answer the speaker? Answer those two and you have your model. This guide puts all three in one table, gives you the decision rule, shows where the Sync and Dictation APIs fit (they stay on Universal-3.6 Pro), and lays out what changes on December 8, 2026 and February 1, 2027 if you're on an older model.
If you're still choosing between async, sync and streaming as an API surface, start with our async, sync, or realtime workshop recap. This post assumes you know which surface you need and want the right model on it. For the launch details and benchmark charts, read the AssemblyAI-STT-3.7 launch post.
The three AssemblyAI-STT-3.7 models at a glance
All three models cover 32 languages and share one API key. They differ in the API they run on, what they're tuned for, which features they carry and what they cost.
| AssemblyAI-STT-3.7 | AssemblyAI-STT-3.7-Realtime | AssemblyAI-STT-3.7-Realtime-Agent | |
|---|---|---|---|
| Use case | Recorded audio: call recordings, finished meetings, interviews, media, medical recordings | Live audio people read: meeting notetakers, live captions, ambient transcription, call recording | Live audio software answers: voice agents, phone agents, drive-thru and kiosk agents |
| API | Pre-recorded (async) | Streaming (WebSocket) | Streaming (WebSocket) |
| Model ID | speech_models: ["assemblyai-stt-3-7"] |
speech_model: "assemblyai-stt-3-7-realtime" |
speech_model: "assemblyai-stt-3-7-realtime-agent" |
| Price | $0.21/hr of audio | $0.30/hr of session time | $0.45/hr of session time |
| Languages | 32, with automatic language detection and native code switching; 99 if you list universal-2 as a second model |
32 in one stream | 32 in one stream, with native code switching |
| Speaker labels | Speaker diarization on the finished file, +$0.02/hr | Realtime speaker labels, final when each turn ends; the docs' recommended model for streaming diarization, +$0.12/hr | Supported, +$0.12/hr |
| Voice Focus | Streaming feature, not applicable | Not available | Yes, near-field or far-field, +$0.10/hr |
| Turn detection | Not applicable | Segments the transcript; the end-of-turn check never runs before 320 ms of silence | Built for turn-taking; manual turn mode (mode: "manual" plus ForceEndpoint) is available on this model only |
| Medical Mode | Yes, +$0.15/hr | Yes, +$0.15/hr | Yes, +$0.15/hr |
| Translation | Not supported. Speech Understanding translation runs on Universal-3.5 Pro and Universal-2 only | Real-time translation of each final turn on the same WebSocket (llm_gateway), billed as LLM Gateway usage |
Real-time translation of each final turn on the same WebSocket (llm_gateway), billed as LLM Gateway usage |
| Keyterms prompting | Up to 1,000 words or phrases, $0.05/hr | Up to 100, included | Up to 100, included |
| Replaces | Universal-3.5 Pro and Universal-3 Pro (pre-recorded) | Universal-Streaming English and Universal-Streaming Multilingual | Universal-3.5 Pro Realtime; successor to Universal-3.6 Pro Realtime, which stays fully supported |
Prices are pay-as-you-go list prices from the AssemblyAI pricing page on November 17, 2026. Streaming is billed on how long the WebSocket stays open, so close sessions when the audio ends. Medical Mode improves medication names, generic and brand, in English, Spanish, German and French.
The decision rule: recorded, read live, or answered live
Ask the questions in this order. The first "yes" picks the model.
- Is the audio already recorded? Use AssemblyAI-STT-3.7 on the pre-recorded Speech-to-Text API. Nobody is waiting on each word, and the model sees the whole file before it writes the transcript.
- Is it live, and do people read the transcript? Use AssemblyAI-STT-3.7-Realtime. A notetaker, a caption bar or an ambient scribe needs words and speaker labels on screen as people talk. Nothing on your side waits for the model to decide the speaker is done.
- Is it live, and does software reply to the speaker? Use AssemblyAI-STT-3.7-Realtime-Agent. Once an agent has to decide when to talk, turn detection, short replies like "no" and "yep," and the digits in an account number decide whether the call works.
Two edge cases come up often:
- Agent assist, where a human agent is on the call and software only suggests. People read the transcript and nobody's software speaks to the caller, so that's Realtime.
- A meeting bot that also answers questions out loud. Once it talks back, it's an agent. Use Realtime-Agent for the stream that drives replies.
Many products run two of these. A notetaker streams through AssemblyAI-STT-3.7-Realtime during the meeting and then sends the recording to AssemblyAI-STT-3.7 for the stored transcript and summary. A contact center runs Realtime-Agent on the phone line and AssemblyAI-STT-3.7 on the recorded calls for QA, both on the same API key.
One API key covers pre-recorded and both realtime models. Sign up free and send your own audio to the model the decision rule picked.
What each model does better than the other two
AssemblyAI-STT-3.7: recorded audio
The pre-recorded model gets the whole file before it writes a word, so it's the one to use for any transcript you store, search or analyze. It covers 32 languages with automatic language detection and native code switching, takes contextual prompting and up to 1,000 keyterms, and runs on a new speaker diarization model built to hold up on short clips. On audio that switches between English and another language, it has the lowest average error rate across the 16 datasets AssemblyAI tests and the lowest on 12 of them, ahead of ElevenLabs, Soniox, Azure, AWS, OpenAI and Deepgram.
Two behaviors to know before you switch. It keeps disfluencies by default (fillers, repetitions, false starts), and with language_detection: true it never falls back to another model unless you list universal-2 after it. Both are in the migration section below.
AssemblyAI-STT-3.7-Realtime: live transcripts people read
This is the streaming model for notetakers, captions and call recording. Speaker labels are final the moment a turn ends, with no revision pass to wait for, so a live view can show "who said what" as the meeting runs. On short-form English its word error rate is 3.23%, ahead of ElevenLabs Scribe v3 Live, AWS Transcribe Streaming and Deepgram Flux in AssemblyAI's launch measurements.
It costs $0.30/hr, against $0.45/hr for the agent model. What it leaves out is turn-taking: per the docs, it detects turns to segment the transcript, and it doesn't carry Voice Focus or manual turn mode. Anything that waits on the model to decide the speaker is done belongs on Realtime-Agent. The streaming speaker diarization guide covers how live labels work in more depth.
AssemblyAI-STT-3.7-Realtime-Agent: live audio software answers
This is the model for anything that answers a caller. It carries everything a voice agent needs from the listening layer: Voice Focus to suppress background voices and room noise, keyterm prompting, context carryover and conversation memory, and manual turn mode if your orchestrator owns endpointing (for example, an external turn detector or a push-to-talk key). The docs call it "the most accurate model for voice agents."
Against Universal-3.6 Pro Realtime on 12,460 English voice-agent calls, number errors on Voice Code Bench fall 80%, and only 1.0% of words change between partial and final, down from 3.2%. Fewer partial rewrites means the text your agent acts on early is the text it gets at the end. The full methodology is in the Realtime-Agent research post. If you're comparing vendors for the agent role, our seven tests for voice-agent speech-to-text work on any provider.
Why the prices differ
The price follows the job. Pre-recorded audio is processed once, as a file, at $0.21 per hour of audio. Both realtime models are billed per hour the session is open. Realtime-Agent, at $0.45/hr, carries the agent features (Voice Focus, manual turn mode, turn detection tuned for replies). Realtime, at $0.30/hr, leaves them out. If nothing in your product talks back, you don't need to pay for the agent features.
The three request shapes in Python
Here's the decision rule and the three configurations in one file, using the Python SDK (pip install assemblyai, version 1.10.11 at the time of writing). The companion repo adds offline tests.
import os
from assemblyai.prerecorded.v2 import Transcriber, TranscriptionConfig
from assemblyai.streaming.v3 import RealTimeParameters
PRERECORDED = "assemblyai-stt-3-7"
REALTIME = "assemblyai-stt-3-7-realtime"
REALTIME_AGENT = "assemblyai-stt-3-7-realtime-agent"
def choose_model(audio_is_live: bool, software_replies: bool) -> str:
"""Return the model ID for a workload."""
if not audio_is_live:
return PRERECORDED
return REALTIME_AGENT if software_replies else REALTIME
def prerecorded_config(speaker_labels: bool = True) -> TranscriptionConfig:
# Add "universal-2" after the 3.7 ID to cover languages outside the 32.
return TranscriptionConfig(
speech_models=[PRERECORDED, "universal-2"],
language_detection=True,
speaker_labels=speaker_labels,
)
def notetaker_params() -> RealTimeParameters:
# Live transcript people read, with realtime speaker labels.
return RealTimeParameters(
sample_rate=16000,
speech_model=REALTIME,
speaker_labels=True,
)
def voice_agent_params() -> RealTimeParameters:
# Live audio an agent answers. Voice Focus is available on this model only.
return RealTimeParameters(
sample_rate=16000,
speech_model=REALTIME_AGENT,
voice_focus="near-field",
)
if __name__ == "__main__":
print("Recorded call ->", choose_model(audio_is_live=False, software_replies=False))
print("Live captions ->", choose_model(audio_is_live=True, software_replies=False))
print("Voice agent ->", choose_model(audio_is_live=True, software_replies=True))
if os.environ.get("ASSEMBLYAI_API_KEY"):
transcriber = Transcriber(
api_key=os.environ["ASSEMBLYAI_API_KEY"], config=prerecorded_config()
)
transcript = transcriber.transcribe("https://assembly.ai/wildfires.mp3")
for utterance in transcript.utterances or []:
print(f"Speaker {utterance.speaker}: {utterance.text}")The two RealTimeParameters objects go into RealTimeTranscriber.connect(); the streaming quickstart shows the full event loop. The WebSocket endpoint is the same for both realtime models. Only speech_model changes.
Upload a recording or stream your microphone in the playground and compare the pre-recorded and realtime models side by side before you write any code.
Where the Sync and Dictation APIs fit
Two more AssemblyAI APIs turn speech into text, and neither runs on AssemblyAI-STT-3.7. Both use Universal-3.6 Pro.
| Sync API | Dictation API | |
|---|---|---|
| Model | Universal-3.6 Pro (model: "universal-3-6-pro") |
Universal-3.6 Pro, plus a hosted rewrite step |
| What you get back | The transcript, in the HTTP response | The verbatim transcript and a cleaned-up version, in one response |
| Audio per request | Up to 120 seconds of WAV or PCM through the SDKs; up to 20 minutes, including compressed formats, over HTTP | Up to 120 seconds through the SDKs; up to 20 minutes over HTTP |
| Price | $0.30/hr | $0.62/hr, rewrite included |
| Use it when | A short clip is finished and someone is waiting, and you handle turn detection yourself | A user speaks into a text field and wants send-ready text |
Set the Sync model explicitly. The Sync docs note that the SDKs still default to universal-3-5-pro, which is retired for Sync, so those requests are routed to universal-3-6-pro anyway. Speaker labels and Medical Mode aren't available on the Sync API. If you need either on short clips, use the pre-recorded API with AssemblyAI-STT-3.7. For the trade-off between the two, read Sync API or Dictation API?.
Migrating from Universal models: December 8 and February 1
AssemblyAI deprecated the previous models on November 17, 2026. Here's what that means for each one, per the docs' model versions and deprecations page.
| If you use | Status | What happens | Move to |
|---|---|---|---|
universal-3-5-pro or universal-3-pro (pre-recorded) |
Deprecated Nov 17, 2026; retirement date to be announced | If you don't set speech_models and run on one of these by default, requests use AssemblyAI-STT-3.7 from December 8, 2026. If you pin the model, nothing changes until it retires. |
assemblyai-stt-3-7 |
universal-2 (pre-recorded) |
GA | Accounts that default to Universal-2 aren't moved automatically | ["assemblyai-stt-3-7", "universal-2"] when you're ready |
universal-3-5-pro (streaming) |
Deprecated Nov 17, 2026; tentative retirement Feb 1, 2027 | After retirement, requests are routed to the replacement | assemblyai-stt-3-7-realtime-agent |
universal-streaming-english or universal-streaming-multilingual |
Deprecated Nov 17, 2026; tentative retirement Feb 1, 2027 | After retirement, requests are routed to the replacement | assemblyai-stt-3-7-realtime |
universal-3-6-pro (streaming) |
GA, fully supported | No change | Realtime-Agent for agents, Realtime for notetakers and captions |
universal-3-6-pro (Sync) |
GA, the only Sync model | No change | Stay on it |
New accounts created on or after November 17, 2026 already default to AssemblyAI-STT-3.7 on pre-recorded requests. AssemblyAI gives at least 90 days' notice before retiring a model.
Before the December 8 switch reaches you, check three behaviors from the AssemblyAI-STT-3.7 migration guide:
- Disfluencies are on by default. AssemblyAI-STT-3.7 keeps fillers, repetitions, stutters and false starts. Set
disfluencies: falseif downstream code expects the cleaner Universal-3.5 Pro output. - No silent fallback with language detection. With
language_detection: trueand onlyassemblyai-stt-3-7listed, a detected language outside the 32 isn't sent to Universal-2. Listuniversal-2second to keep the old fallback. - Use
language_codesto restrict languages.language_detection_options.expected_languagesand thecode_switchingoptions apply only to Universal-2.
On streaming, the change is one string: swap speech_model to the ID in the table. The Realtime Speech-to-Text API page lists both 3.7 models, and the endpoint is unchanged.
Choose by the job, then by the price
Most model mistakes come from picking on price first: streaming a recording through a realtime model, or running a notetaker on the agent model because it was the newest. The 3.7 lineup makes the job-first choice easy. Recorded audio goes to AssemblyAI-STT-3.7. Live audio people read goes to AssemblyAI-STT-3.7-Realtime. Live audio your software answers goes to AssemblyAI-STT-3.7-Realtime-Agent. Sync and Dictation stay on Universal-3.6 Pro for finished short clips and voice input.
The split also means you can mix models inside one product without a second vendor. Teams building voice agents often add a recorded-call pipeline later, and that second job gets its own model on the same key. If you're comparing AssemblyAI with other vendors before you get here, our guide to choosing a speech-to-text API walks through the criteria. If your product adds translated captions, the real-time speech translation guide shows the llm_gateway pattern on the realtime models.
Pre-recorded, realtime and realtime agent models on one account with pay-as-you-go pricing. Create a free account and make your first request in minutes.
Frequently Asked Questions (FAQs)
Which AssemblyAI model should I use for a voice agent?
Use AssemblyAI-STT-3.7-Realtime-Agent (assemblyai-stt-3-7-realtime-agent) for any application where software replies to the speaker. It's the only 3.7 model with Voice Focus and manual turn mode, and its turn detection is built for conversations. It costs $0.45 per hour of session time.
What's the difference between AssemblyAI-STT-3.7-Realtime and AssemblyAI-STT-3.7-Realtime-Agent?
AssemblyAI-STT-3.7-Realtime is built for live transcripts that people read, such as notetakers and captions, with realtime speaker labels at $0.30/hr. AssemblyAI-STT-3.7-Realtime-Agent is built for voice agents, adding Voice Focus and manual turn mode at $0.45/hr. Both cover 32 languages on the same streaming endpoint.
Is the Sync API on AssemblyAI-STT-3.7?
No. The Sync API runs on Universal-3.6 Pro (universal-3-6-pro), which is its only model. Set it explicitly in your config, because the SDKs still default to a retired Sync model ID that gets routed to Universal-3.6 Pro.
What happens to Universal-3.5 Pro on December 8, 2026?
From December 8, 2026, existing accounts that don't set speech_models and use Universal-3.5 Pro or Universal-3 Pro by default automatically move to AssemblyAI-STT-3.7 on pre-recorded requests. The switch changes only the default model: Universal-3.5 Pro isn't retired yet, and requests that pin it keep working. Universal-3.5 Pro Realtime and the Universal-Streaming models have a tentative retirement date of February 1, 2027.
How much does AssemblyAI cost per hour?
At pay-as-you-go list prices, AssemblyAI-STT-3.7 costs $0.21 per hour of audio, AssemblyAI-STT-3.7-Realtime costs $0.30 per hour of session time and AssemblyAI-STT-3.7-Realtime-Agent costs $0.45 per hour of session time. The Sync API is $0.30/hr and the Dictation API is $0.62/hr. Add-ons such as speaker labels and Medical Mode are priced separately.
Which speech-to-text model is best for meeting notetakers?
For a live notetaker, AssemblyAI's docs recommend AssemblyAI-STT-3.7-Realtime, which returns speaker labels during the meeting. For the stored transcript after the meeting ends, send the recording to AssemblyAI-STT-3.7 on the pre-recorded API. Many notetakers run both.