Gradium's stt-translate and s2s-translate beat GPT on speech translation latency
Gradium released two real-time speech translation models that collapse the standard three-model cascade (ASR + MT + TTS) into a single duplex WebSocket, cutting latency and improving accuracy versus gpt-realtime-translate and gemini-3.5-live-translate.
• Covers 20 language pairs: English, French, German, Spanish, Portuguese
• Single-pass transcription-and-translation paired with Gradium TTS
• Voice selection and cloning on output