mapleeve/voscript repository overview

⁠VoScript

Meeting recordings → timestamped transcripts with real speaker names. Self-hosted, GPU-powered, remembers every voice.

You have a meeting recording with six people. You want to know who said what. Whisper gives you a wall of text. pyannote splits it into "Speaker A / B / C" — but it doesn't know who anyone is. You still label every recording by hand.

VoScript fixes that: enroll a voice once, and it gets auto-identified in every future recording. Not "Speaker 2" — "Maple".

Audio ──► faster-whisper large-v3    transcription + word-level timestamps
      ──► pyannote 3.1               speaker diarization
      ──► WeSpeaker ResNet34          speaker embeddings
      ──► VoiceprintDB (AS-norm)      match against enrolled voices
      ──► transcript with real names

⁠Quick start

git clone https://github.com/MapleEve/voscript.git && cd voscript
cp .env.example .env   # set HF_TOKEN and API_KEY
docker compose up -d --build
curl -sf http://localhost:8780/healthz

⁠What's inside

• Persistent voiceprint library — enroll once, auto-match across all recordings. sqlite + sqlite-vec, scales to thousands of speakers
• AS-norm scoring — impostor cohort built at startup; eliminates per-speaker baseline bias
• Adaptive threshold — each speaker's threshold relaxes based on enrollment variance
• Speaker cluster consolidation — diarization splits merged to a single label automatically
• Word-level timestamps — WhisperX forced alignment
• Optional denoising — DeepFilterNet / noisereduce with SNR gate (clean audio skips denoising)
• File dedup — same file submitted twice returns existing result immediately
• Plain HTTP API — multipart/form-data, no SDK required

⁠Tags

⁠Environment variables

⁠Links

• GitHub⁠
• API Reference⁠
• Quickstart⁠
• Changelog⁠