AI Models

AssemblyAI Launches STT-3.7 Speech Models for Agents

November 17, 2026 · By Editorial Team

AssemblyAI Launches STT-3.7 Speech Models for Agents

AssemblyAI has launched STT-3.7, a new generation of speech-to-text models that splits transcription into three specialised versions for different jobs. The release, which began rolling out on November 17, 2026, replaces the company’s previous Universal models as its recommended speech layer.

The three models are AssemblyAI-STT-3.7 for pre-recorded audio, STT-3.7-Realtime for live meetings and notetakers, and STT-3.7-Realtime-Agent for voice-agent conversations. All three support 32 languages with automatic detection, and the realtime models add speaker labels, keyterm prompting, context carryover and conversation memory.

The split reflects how differently each use case listens. A meeting notetaker needs clean, final speaker labels at the end of each turn, while a voice agent must decide when a caller has finished speaking and filter out background chatter. AssemblyAI says its own benchmarks show number errors down 80% and background chatter roughly halved, though those are vendor figures and independent tests have yet to appear.

Live transcription is priced from $0.30 per session hour, and the voice-agent model from $0.45. The launch lands in a crowded speech market. Deepgram remains a leading alternative for realtime transcription, while Otter.ai and Fireflies.ai package transcription into meeting products rather than selling the raw API. AssemblyAI’s bet is that developers building voice agents will pay for a speech layer tuned to their exact workload rather than a one-size-fits-all model.

Source: AssemblyAI Docs

Related stories