Gradium builds voice models for real-time agents and exposes them through one API: streaming text-to-speech, streaming speech-to-text, live voice-to-voice translation, and voice cloning. The models target the turn-by-turn loop of a voice agent, so the speech-to-text side includes semantic voice activity detection for turn-taking, a flush control, and adjustable delay for trading latency against context.
Text-to-speech streams over a WebSocket and returns word-level timestamps for aligning text and audio. One model handles English, French, German, Spanish, and Portuguese while keeping the same speaker identity across languages. Voices come from a voice library or from cloning: an instant clone from about 10 seconds of audio, or a pro clone fine-tuned for higher speaker similarity. Speed, temperature, and similarity are adjustable per request. Gradium reports a time-to-first-audio of 158 ms (P50) with a 2 ms latency interquartile range on the Coval benchmark.
The API is served from EU and US clusters with automatic regional routing at api.gradium.ai, authenticated with an x-api-key header, and there is a Python SDK, example voice-agent repositories, and a Pipecat TTS service. An on-device text-to-speech model runs offline, and Gradbot builds a voice agent from a single prompt.
Features
- Streaming text-to-speech: WebSocket synthesis with word-level timestamps and adjustable speed, temperature, and similarity
- Streaming speech-to-text: real-time transcription with semantic VAD, flush, and adaptive delay controls for agent turn-taking
- Live translation: voice-to-voice translation in real time
- Voice cloning: instant clones from roughly 10 seconds of audio and pro clones fine-tuned for speaker similarity
- Multilingual single model: English, French, German, Spanish, and Portuguese with a consistent speaker identity
- Regional routing: EU and US clusters selected automatically for data residency
- On-device TTS: an offline real-time model for local synthesis
- Developer tooling: Python SDK, REST and WebSocket references, a Pipecat service, and voice-agent examples
- Gradbot: generates a working voice agent from one prompt
