Two new models, both on ONNX Runtime and burn.
Multitalker TDT (MultitalkerTDT, feature sortformer)
I fine tuned parakeet-tdt-0.6b-v3 fine tuned with NVIDIA's speaker kernels. With Nemotron-3 Diarization it gives one transcript per speaker, also when people talk over each other, in all 25 languages. With 2 overlapping speakers, cpWER goes from 36.7% to 4.4% on LibriSpeechMix and from 39.0% to 10.7% on six other languages (MLS mixtures), compared to plain TDT plus diarization. Single speakers stay as good as plain TDT.
Model: multitalker-tdt-0.6b-v3, details in the blog post.
Unified, 25 languages (ParakeetUnified)
NVIDIA's new nemotron-labs-unified-asr-0.6b: offline and buffered streaming in one model. Same API as the English Unified, plus
set_target_lang("de") (default auto).
Model: nemotron-labs-unified-asr-0.6b
Also
- Multitalker now picks fp32 when a folder has both fp32 and int8, like the other models. Keep only the int8 files if you want int8.
What's Changed
- Multitalker TDT: per-speaker transcripts in 25 languages, overlaps included by @altunenes in #146
- Burn multitalker tdt encoder by @altunenes in #147
- Unified multi by @altunenes in #148
- multitalker: prefer fp32 like the other models by @altunenes in #149
Full Changelog: v0.4.3...v0.4.4