AssemblyAI makes its voice-agent speech model generally available

AssemblyAI announced general availability on October 27, 2026, giving voice-agent developers a text-to-speech model for spoken replies. Its listed price is $15 per million input characters, including spaces and punctuation. The six launch languages are English, Spanish, French, German, Italian and Portuguese; Hungarian is not listed.

The WebSocket specification describes a persistent connection that accepts text as it arrives. Voice and sample rate are chosen when connecting; changing them requires a new session. Developers must wait for the server’s Begin message before sending text: an open socket alone does not establish successful authentication, because rejected credentials produce an error afterwards. The client should use the effective configuration returned in Begin, rather than assume the requested settings took effect.

The OpenAI-compatible route lets an OpenAI SDK use AssemblyAI’s base address and API key. The request needs input text and an AssemblyAI model name; an OpenAI model name is rejected. Compatibility also has boundaries: OpenAI voice names map to AssemblyAI voices, while the instructions field is ignored. Audio streams as it is synthesized, and a failure after streaming begins cuts the connection.

MarkTechPost also covered the release.