Local Speech to Text: Whisper

Local speech to text in 2026: Whisper for accuracy, faster-whisper for speed, FunASR for Chinese, with model sizes and install code.

Local speech to text without the cloud

Three projects cover the space. openai/whisper (111,299 stars) is the accuracy baseline with 99 languages. SYSTRAN/faster-whisper (26,163 stars) reimplements the same weights on CTranslate2 for several times the speed and less VRAM. modelscope/FunASR (20,824 stars) targets Chinese and outputs punctuation and timestamps out of the box. whisperX (24,855 stars) adds forced alignment and speaker diarization. Star counts as of October 2026.

Model sizes: tiny 39M, base 74M, small 244M, medium 769M, large-v3 1550M, distil-large-v3 756M. Rough VRAM at fp16 runs from under 1GB for tiny to about 10GB for large-v3. CPU-only works but is much slower.

Practical gain comes from three things more than from the model: enable VAD to skip silence, pass a terminology list through initial_prompt (Whisper) or hotword (FunASR), and keep segments at 10 to 20 seconds.

Pick Whisper for the fastest start, faster-whisper for long recordings on a GPU, FunASR for Chinese meetings, WhisperX when you need speaker labels. Licenses are MIT and Apache 2.0, verify on each repository.