altunenes/parakeet-rs v0.2.4

v0.2.4 - Sortformer v2 Streaming Speaker Diarization

on GitHub

What's New

NVIDIA Sortformer v2 - Complete rewrite of speaker diarization with streaming support. Handles long audio files (25+ minutes) natively without memory issues.

Features

• Streaming Architecture: Processes audio in ~10s chunks with speaker cache for consistent speaker tracking across entire recordings
• Pre-tuned Configs: Choose from callhome(), dihard3() or custom() (your custom use case)
• Custom Thresholds: Fine-tune onset/offset thresholds and minimum segment durations for your use case
• 4 Speaker Support: Identifies and tracks up to 4 distinct speakers

Usage

  use parakeet_rs::sortformer::{Sortformer, DiarizationConfig};

  let mut sortformer = Sortformer::with_config(
      "diar_streaming_sortformer_4spk-v2.onnx",
      None,
      DiarizationConfig::callhome(),
  )?;

  let segments = sortformer.diarize(audio, 16000, 1)?;

Download

Get the model: https://huggingface.co/altunenes/parakeet-rs/blob/main/diar_streaming_sortformer_4spk-v2.onnx

See examples/diarization.rs for combining diarization with TDT transcription.

Full Changelog: v0.2.3...v0.2.4

add diar_streaming_sortformer_4spk-v2 onnx by @altunenes in #23