Text-to-Speech and Speech-to-Text
Last updated 21 Oct 2026
TTS providers
ElevenLabs, OpenAI tts-1 family, Cartesia, PlayHT, plus our compute-fund hosted XTTS for batch jobs. Pricing is PerSecondAudio (modern) or PerCharacterTts (legacy) — both are tracked. Voice cloning requires a recorded consent sample; you must hold the rights to any voice you clone.
STT providers
Whisper (OpenAI), Deepgram, AssemblyAI, Speechmatics. Diarization, word-level timestamps, and language detection are first-class capabilities; pick the provider that supports the ones you need via the capability filter.
Realtime voice
VoiceSession entities track full-duplex conversations (input → model → output streamed). Realtime models (gpt-4o-realtime, Gemini Live, etc.) bill PerMinuteRealtime rather than per-token. Latency targets: <300ms first audio, <500ms steady state.
Languages supported
TTS and STT cover the major world languages and many more. Quality varies by language and provider: the most widely spoken languages get the most natural voices, and transcription accuracy follows the training data available for each language.
The supported language list is in /docs/doc-api-reference#audio-languages.
Watch how it works
Live translation, live transcription and translated video subtitles with AI
Choose the language you speak and the one to translate into, and start talking: your voice is translated as you go. It works one direction at a time; tap swap and the other person answers in their own language, and the whole conversation can be saved as a note. Live Transcription writes down what you say in real time, Speech-to-Text turns a recording into text, and a video comes back with its transcript, its translation and subtitle files in both languages.