New AssemblyAI-TTS-1: Text-to-speech for voice agents. Learn more
Speaker diarization

Speaker diarization accurate enough for production

Word-level speaker labels for pre-recorded and realtime audio, from our most accurate speaker diarization yet—models AssemblyAI builds and trains in-house.

Word-level attribution

Every word carries a speaker label, even through overlap and interruptions, so no misattributed words corrupt your transcripts.

Streaming and async

Label speakers in real time on the Realtime Speech-to-Text API or in batch on the Pre-recorded API, with diarization models AssemblyAI builds and trains in-house.

10+ speakers, 99 languages

Consistent speaker labels for 10+ speakers on Universal-3.5 Pro, Universal-2 (99 languages), and Universal-3.6 Pro Realtime, from a two-person call to a full meeting.

Metaview
Ashby
Cluely
Genio
Siro

36%

improvement in close rate

See case study
LiveKit
EliseAI
Commure
Dovetail
Fireflies

“The new Universal-3.5 Pro speech model from AssemblyAI is best so far in terms of accuracy, latency, and language switching.”

Retell
CallRail
Apollo.io
ClickUp
Calabrio

80%

increase in customer satisfaction

HeyGen
Granola
Siro
JotPsych
Granola

“Assembly has saved us countless hours managing models, and provided exceptional accuracy.”

Metaview
Ashby
Cluely
Genio
Siro

36%

improvement in close rate

See case study
LiveKit
EliseAI
Commure
Dovetail
Fireflies

“The new Universal-3.5 Pro speech model from AssemblyAI is best so far in terms of accuracy, latency, and language switching.”

Retell
CallRail
Apollo.io
ClickUp
Calabrio

80%

increase in customer satisfaction

HeyGen
Granola
Siro
JotPsych
Granola

“Assembly has saved us countless hours managing models, and provided exceptional accuracy.”

Quickstart

See who said what, minutes after you sign up

Create a free account and add speaker labels to any audio with one flag. Test it in the no-code playground, then copy a ready-made request into your app. Speaker diarization is $0.02/hr on pre-recorded audio and $0.12/hr on the Realtime Speech-to-Text API.

Start building free

No credit card required

Benchmarked on what your users actually read

The industry-standard DER compares time regions, never words—a perfect transcript can score a “broken” DER on formatting alone. cpWER measures who-said-what accuracy the way your users judge it: for each speaker, what fraction of their words did the system get wrong or credit to someone else?

Read why we measure with cpWER

Average cpWER across DiPCo, CALLHOME, NOTSOFAR and AMI.

*Lower is better*
Speaker-attributed word error rate (cpWER)
AssemblyAI Universal-3.5 Pro
30.17%
Azure
30.35%
ElevenLabs Scribe V2
35.26%
Speechmatics
36.60%
Gladia
36.88%
Mistral Voxtral
37.52%
Deepgram
37.93%

Source: AssemblyAI published benchmarks (assemblyai.com/benchmarks), pre-recorded tab, updated July 2026.