Pinch

Pinch translates speech into other languages for voice agents, live calls, and recorded media. Two services cover the two timescales: a real-time translation API that streams microphone or call audio over a WebSocket and returns translated audio plus transcripts, and an asynchronous dubbing API that takes audio or video files and returns a dubbed version in another language.

The real-time API runs in speech-to-speech or speech-to-translated-text mode and needs no SDK. It accepts 32 input languages (ISO-639-1 codes, regional variants accepted); 30 of them also have synthesized voice output, while Filipino and Cantonese return text only. Passing auto as the source language turns on language detection. The relay-1 translation model emits transcription, translated transcription, and speech output at the same time and is offered via API and on-prem deployment.

Dubbing keeps the original speaker's voice and tone, phrases the translation the way a native speaker would rather than word for word, and isolates speech so music, ambience, and sound design pass through unchanged. A browser studio lets you check speaker identification, reassign or merge speakers, and edit the transcript and translation before generating the dub. Pinch also publishes research on JEPA-based audio encoders for voice-preserving speech-to-speech translation.