AssemblyAI introduces STT-3.7 models for live conversations, agents and recordings

AssemblyAI introduced its STT-3.7 transcription family, with separate models for live conversations, voice agents and recorded audio, in an announcement now dated November 18.
All three support 32 languages with automatic detection. The Realtime model targets meetings, calls and captions, with speaker labels finalized when a speaking turn ends. Realtime-Agent is tailored to conversations with AI agents, including short replies, numbers and competing background speech.
Listed prices are $0.30 per audio hour for Realtime, $0.45 for Realtime-Agent and $0.21 for recorded audio. Developers select the appropriate model through the existing realtime or pre-recorded speech-to-text APIs.
The recorded-audio release adds 14 languages to the previous model’s 18 and supports speakers switching languages within a conversation. Coverage beyond those 32 languages requires explicitly adding Universal-2 as a fallback; it is not automatic.
AssemblyAI reports accuracy and speaker-labeling improvements, but those claims are not independent testing by Guth News. The practical change is a choice of transcription models optimized for distinct workflows, rather than treating meeting captions and conversational agents as the same task.
Sources and citations
Revision history
Published .