FunASR

FunASR is the open-source speech toolkit from Alibaba's ModelScope team and one of the most production-credible alternatives to OpenAI Whisper in 2026. It bundles a family of in-house models — Paraformer for non-autoregressive ASR, SenseVoice for multilingual recognition with emotion and event detection, CAM++ for speaker verification, and FSMN-VAD for voice activity detection — into a single toolkit with a unified Python API and a self-hostable HTTP server. Headline numbers are aggressive: 170x realtime decoding on a modern GPU, 50+ languages, robust performance on Chinese and other Asian languages where Whisper has historically struggled, and built-in speaker diarisation, timestamping, punctuation and streaming. The server speaks an OpenAI-compatible transcription API, so teams can swap it in behind existing Whisper integrations with no client changes. FunASR has become the default ASR backbone for many Chinese-language voice agent stacks and is increasingly used worldwide by teams who want on-prem speech without paying per-minute cloud rates. It is Apache-licensed, ships pre-built Docker images for CPU and GPU inference, and integrates cleanly with the ModelScope hub for newer model releases.