altunenes/parakeet-rs v0.3.8

on GitHub

Highlights

Nemotron 3 Diarization (Sortformer v3) replaces Sortformer v2: up to 8 speakers, much better on overlapping speech, and all four of NVIDIA's latency presets (30.4 s to 0.32 s) from a single ONNX via StreamingProfile. Validated against NVIDIA NeMo's output. Model: nemotron3_diar_v3.onnx

Breaking changes

• Sortformer v2 models are no longer supported. Download nemotron3_diar_v3.onnx; loading a v2 ONNX returns an error.
• NUM_SPEAKERS is now 8.
• DiarizationConfig::callhome() and dihard3() are removed (they were tuned for v2). DiarizationConfig::default() now matches NeMo: 0.5 threshold, no
  smoothing.
• Raw predictions (predict_raw, diarize_chunk_raw) now have one row per 10 ms (was 80 ms). median_window counts 10 ms frames.
• The default streaming latency is NVIDIA's offline preset (30.4 s buffer). For live audio, use set_profile(StreamingProfile::low_latency()).
• ModelConfig::with_custom_configure closures must now be Send + Sync (#120).

What's Changed

• fix: use TDT durations for timestamps by @nemosupremo in #118
• Make ModelConfig::build_session public by @rbanderovskyi in #119
• Make ModelConfig Send + Sync by @rbanderovskyi in #120
• Allow a separate execution config for the TDT decoder/joint by @rbanderovskyi in #121
• Split diarize into predict_raw and post_process by @rbanderovskyi in #125
• Remove the now-unused ParakeetTDTModel::from_pretrained by @rbanderovskyi in #124
• Add a per-window session router to Sortformer by @rbanderovskyi in #123
• Trim the explanatory comments from the diarize split by @rbanderovskyi in #126
• Add a verified Orukeet TDT example by @Nathan-Roll1 in #128
• Add sortformer v3 by @altunenes in #129

New Contributors

• @nemosupremo made their first contribution in #118
• @rbanderovskyi made their first contribution in #119
• @Nathan-Roll1 made their first contribution in #128

Full Changelog: v0.3.7...v0.3.8