AssemblyAI launches Universal-2 streaming transcription for voice agents

Read the original article →

What happened

AssemblyAI released Universal-2 on August 2, a streaming speech-to-text model for voice-agent pipelines with 99 language support and sub-250ms partial transcript latency.

The model adds speaker diarization, sentiment detection, and PII redaction in a single API call. AssemblyAI said Universal-2 reduces word error rate by 15 percent compared with Universal-1 on conversational audio. Pricing starts at $0.0065 per minute.

Why it matters

Voice agents need accurate streaming transcription as the foundation layer. Bundling diarization, sentiment, and PII redaction reduces the number of API calls in a voice pipeline.

PII redaction in particular matters for healthcare and financial agents subject to data-handling rules.

MintedBrain take

Compare Universal-2 against Deepgram and Whisper on your accent and domain vocabulary before switching. WER improvements on benchmarks may not translate to your specific audio conditions.

References

This article was originally published at AssemblyAI docs. For the full piece, read the original article.

Discussion

  • Loading…

← Back to News