Skip to main content

JD Transcription

AssemblyAI Releases STT-3.7 With 32 Languages and New Speaker Labels

We track speech recognition releases. AssemblyAI has released AssemblyAI-STT-3.7, a new model for recorded audio. The company lists the model as generally available in its speech to text API for prerecorded files. It replaces the earlier Universal models as the default choice for recorded audio.

AssemblyAI says the model fits meetings, calls, podcasts, media, clinical recordings, and voice notes.

More languages and code switching

The new model supports 32 languages. The previous model covered 18. The new languages include Urdu, Russian, Korean, Catalan, Galician, Romanian, Estonian, Persian, Cantonese, Afrikaans, Marathi, Zulu, Xhosa, and Norwegian Nynorsk.

The model also handles code switching. Code switching happens when a speaker moves between two languages in one conversation, or even in one sentence. AssemblyAI says the model returns each word in the language and script the speaker used. The company reports an 80 to 90 percent drop in language flips. A language flip is a transcript that changes language while the speaker stays in the same language.

AssemblyAI also reports gains in the 17 languages other than English that its previous model covered. Their average WER fell from 12.8 percent to 10.1 percent. WER counts word mismatches against a reference transcript, so a lower number is better.

A new speaker labeling model

Speaker labels, also called diarization, show who said what in a recording. We follow diarization closely because it shapes how useful a transcript is for meetings, calls, and interviews. AssemblyAI shipped a new diarization model with the release. The company is rolling it out across all of its models and languages.

On AssemblyAI’s internal test sets, a score called DER fell from 14.7 percent to 10.4 percent. DER measures the share of audio that carries a mismatched speaker label, so a lower number is better. That change is a drop of about 29 percent. Another score, cpWER, combines word accuracy and speaker accuracy. It fell from 29.3 percent to 24.0 percent. The model also counts speakers more closely. Its average gap on the speaker count fell from 0.57 to 0.20.

Short clips showed the largest gain. On 60 second clips, cpWER fell from 34.0 percent to 19.1 percent. The company says speaker labels now hold up on voice notes and short call snippets.

Pricing and default settings

The base price stays at $0.21 per hour of audio, the same rate as the earlier model. Extra features carry their own prices. The AssemblyAI pricing page lists standard speaker labels at $0.02 per hour and keyterm prompting at $0.05 per hour.

New accounts created on or after November 17, 2026 use the new model by default. Starting December 8, 2026, existing accounts that use the earlier Universal Pro models by default will move to the new model automatically. AssemblyAI says this is a change to the default model. The older models keep running, and the company will announce a retirement date later.

Developers who need languages beyond the 32 can pair the new model with an older AssemblyAI model. That pairing covers 99 languages. We will keep tracking speech recognition releases as new models reach transcription workflows.

Sources