AssemblyAI Universal-3.6 Pro Realtime STT

[tschellenbach]

Why

Universal-3.6 Pro Realtime (universal-3-6-pro, launched 2026-09-29) is first on Coval's streaming STT WER leaderboard (2.2% WER, 138 ms to final segment, October 2026). Artificial Analysis has not benchmarked it yet. It adds 32 languages with code-switching, keyterms and word-aware end-of-turn detection to the STT router.

Facts checked against AssemblyAI's own docs and pricing page, and confirmed on the live socket:

Fact	Confirmed
Price	$0.45/hr, billed on session duration, keyterms included
Auth	Authorization: <api key>, no Bearer prefix (socket_test.go TestTheKeyTravelsAsItIsInTheAuthorizationHeader)
speech_model enum	lists universal-3-6-pro, and Begin echoes it back as configuration.model

How it works

sequenceDiagram
    participant C as Call (20 ms frames)
    participant P as assemblyai.STT
    participant S as streaming.assemblyai.com/v3/ws
    P->>S: dial with speech_model, mode, keyterms_prompt, language_codes and Authorization
    S-->>P: Begin with configuration
    Note over P: refuse if the model or preset differs from what was asked
    C->>P: ProcessAudio 20 ms
    P->>S: binary PCM, gathered to at least 50 ms and split at 1000 ms
    S-->>P: Turn with end_of_turn false
    P-->>C: Transcript ModeReplacement
    S-->>P: Turn with end_of_turn true
    P-->>C: Transcript ModeFinal
    C->>P: Close
    P->>S: tail padded to 50 ms, then Terminate
    S-->>P: last Turn, Termination, close 1000

Loading

Change	Result	Proven by
Begin check (assemblyai.go:351)	unknown params are silently ignored, so a misspelled model or preset fails at Start	TestAMisspeltModelIsRefusedRatherThanServedByTheDefault (live), TestASessionOpenedOnAnotherModelIsRefused
Frame gathering (assemblyai.go:389)	the server rejects frames under 50 ms with 3007	TestCallAudioIsGatheredIntoFramesTheServerAccepts, live suite at ChunkMs: 20
Default mode=max_accuracy	balanced ended a turn at the 390 ms pause after "a map." in the fixture	TestACallGetsThePresetThatDoesNotEndATurnMidSentence, live TestWordsArriveWhileTheCallerIsStillSpeaking
Terminate on Close (assemblyai.go:410)	the tail settles, and billing stops when the call ends	TestClosingSendsTheLastAudioAndAsksForTheTurnToSettle, live TestClosingSettlesTheTailOfTheCall
Registry and router.yaml	silence_ms sets max_turn_silence. overwrites sets mode, min_turn_silence and max_turn_silence	TestAssemblyAITakesItsPresetFromOverwritesAndRefusesOneItHasNot
sttsuite clock test	AssemblyAI writes "party of 4" as a digit	live TestAMidUtteranceClockTimeSettlesAsOneTranscript

What the router config declares, per router-stt:

Supports	Refuses (not declared)
keyterms (100 terms, 50 characters each), endpointing, realtime, 32 languages	diarize, language detection, redaction, profanity filter. The model can do these, but nothing here sends them

The data policy is trains_on_data: "yes" and retention: unspecified, because self-serve accounts are in the model-improvement program until an admin opts out.

Live suite: all 8 tests pass against the real API, and the default turn settles about 760 ms after the last word. Free-tier accounts allow about 5 new sessions per minute. One earlier run, started straight after a burst of probe sessions, failed two tests in 60 ms each. That fits the rate limit, but I did not capture the error message, so it is unverified.

Changes

• New package acceleration/internal/stt/assemblyai, with unit tests, fake-socket tests and the integration suite
• Registry entry in sttrouter, plus a router.yaml entry for assemblyai/universal-3-6-pro
• sttsuite accepts a digit for "four"
• CHANGELOG.md entry