Model files

Model files for download, attached to the release models-1.

None of these models is ours. Each one is the work of its authors, listed below, and is shared here unchanged or converted, under its own license. Nothing here grants any rights beyond those licenses. Each zip holds an ATTRIBUTION.txt with its source, its license and what was changed.

Sources and licenses

• melllm-zipformer2-66m-he-1.zip: ivrit-ai/zipformer2-66m-streaming-he, by Yakov Kolani (original), via ivrit.ai. CC-BY-4.0. Files unchanged.
• melllm-zipformer2-en-1.zip: csukuangfj/sherpa-onnx-streaming-zipformer-en-2023-06-26, by the k2-fsa project. Apache-2.0. Files renamed only.
• melllm-speakers-1.zip:
  • pyannote/segmentation-3.0, © 2022 CNRS, MIT, in the ONNX export by k2-fsa;
  • NVIDIA NeMo TitaNet-small, Apache-2.0 (the NeMo Toolkit's license), in the ONNX export by k2-fsa.
  • Files renamed only.
• melllm-whisper-turbo-q5k-he-1.zip:
  • ivrit-ai/whisper-large-v3-turbo-ggml, by ivrit.ai (fine-tuned from OpenAI's Whisper large-v3 turbo), Apache-2.0. Quantized to 5 bits (q5_k) with whisper.cpp.
  • Silero VAD, © Silero Team, MIT, as published by k2-fsa. Unchanged.
• melllm-punct-he-1.zip: Horizon-Labs/punctuation-restoration-small, Apache-2.0. Its token table cut to what Hebrew text uses, weights stored in 16 bits, and a smaller tokenizer file.

All files are provided as they are, without warranties of any kind.