WhisperX

Open-source local speech-to-text software that uses Whisper models, adds word-level timestamps and optionally labels speakers.

Alternative to:

WhisperX is a local speech-to-text tool for transcribing interviews, meetings and long recordings. It builds on OpenAI's Whisper, adding individual word timings and optional speaker identification for subtitles and transcript review.

Word timestamps: wav2vec2 alignment places timestamps on individual words, helping you match subtitles to speech or find a passage in a recording.
Speaker labels: Optional pyannote-audio diarization assigns speaker IDs to distinguish voices in a conversation.
Long recordings: The faster-whisper backend supports batched transcription, with voice activity detection to focus processing on speech.
Language and context: Language-specific alignment models support English, French, German, Spanish and Italian by default, with other models available through Hugging Face; optional context between segments can help with technical language and punctuation.

WhisperX uses the BSD 2-Clause license and can run on a CPU or use NVIDIA GPU acceleration through CUDA. The large-v2 transcription setup described by the project needs less than 8 GB of GPU memory, though larger models and settings can increase memory use.

Speaker labeling requires a Hugging Face access token and acceptance of the speaker-diarization-community-1 model agreement. Word alignment requires an appropriate model for the recording's language; languages outside the supplied defaults may need a separate alignment model tested on your audio.