Parafoil/speech-models

Parafoil speech models

The models the Parafoil desktop app downloads to transcribe meetings on the user's own computer, and, in private mode, to remove other people's details from the transcript before it leaves that computer. Each folder holds one model pack. The app pins every file by size and SHA-256 and refuses any file that does not match.

parakeet-tdt-0.6b-v3-pc8-2026-10-15

Speech recognition in 25 European languages, for sherpa-onnx (nemo_transducer).

• encoder.int8.onnx, decoder.onnx, joiner.onnx, tokens.txt are derived from NVIDIA Parakeet TDT 0.6B v3, by NVIDIA, licensed under CC-BY-4.0. The source was the sherpa-onnx ONNX export of that model.
• What we changed: the encoder was quantized with ONNX Runtime dynamic quantization, per-channel int8 weights on MatMul nodes only; convolution layers stay fp32. The decoder, joiner and tokens are unchanged. On AMI meeting audio it scored 13.7% word error rate in our tests.
• silero_vad.onnx is Silero VAD, by the Silero Team, under the MIT licence. Unchanged.

nemotron-3-diarization-q8-2026-09-23

• Nemotron-3-Diarization.q8_0.gguf is NVIDIA Nemotron 3 Diarization, by NVIDIA, under the OpenMDW License Agreement 1.1. It is an unchanged copy of NVIDIA's file at commit f667ed73aee57d40cc39428eb768b4fd87a0a29e. NVIDIA's model card, with its notices, is kept in licenses/Nemotron-3-Diarization-README.md.

gliner-multi-pii-v1-wo8-2026-10-24

The redaction model private mode uses, replacing gliner-multi-pii-v1-pc8e-2026-10-23. It finds names, dates, phone numbers, places, ID numbers and other personal details in a transcript, so private mode can replace them on the device. It runs with ONNX Runtime 1.28 or later.

• model.onnx is derived from GLiNER multi PII v1 (commit 1fcf13e85f4eef5394e1fcd406cf2ca9ea82351d), by Urchade Zaratiana, licensed under the Apache License 2.0. gliner_config.json is that model's configuration, unchanged.
• What we changed: the model was exported to ONNX with the gliner 0.2.29 library and quantized with ONNX Runtime 1.31's tools, weights only:
  • the vocabulary table is per-channel int8, dequantized on load (quantize_dynamic, op_types_to_quantize=["Gather"]);
  • every MatMul is 8-bit MatMulNBits, block size 32, symmetric, accuracy_level=1.
• Why it replaces the earlier pack: this model computes in float32 throughout. The earlier pack quantized activations to int8 at run time, which depends on the processor's int8 instructions. On processors without them (Apple M1, for one), it removed only about half the identifiers. On our synthetic test set this pack removes 97.0% of personal identifiers.
• spm.model and tokenizer.json are the tokenizer of mDeBERTa-v3 base (commit a0484667b22365f84929a935b5e50a51f71f159d), by Microsoft, under the MIT licence. spm.model is unchanged. tokenizer.json was converted from it by the Hugging Face transformers library and adds GLiNER's two tokens, <<ENT>> and <<SEP>>.

gliner-multi-pii-v1-pc8e-2026-10-23

Superseded by gliner-multi-pii-v1-wo8-2026-10-24, and kept unchanged, as every folder here is.

Finds names, dates, phone numbers, places, ID numbers and other personal details in a transcript, so private mode can replace them on the device. It runs with ONNX Runtime 1.28 or later.

• model.onnx is derived from GLiNER multi PII v1 (commit 1fcf13e85f4eef5394e1fcd406cf2ca9ea82351d), by Urchade Zaratiana, licensed under the Apache License 2.0. gliner_config.json is that model's configuration, unchanged.
• What we changed: the model was exported to ONNX with the gliner 0.2.29 library, then quantized with ONNX Runtime 1.31 dynamic quantization: per-channel int8 weights on MatMul and Gather nodes (quantize_dynamic(..., weight_type=QInt8, per_channel=True, op_types_to_quantize=["MatMul", "Gather"])). Plain per-tensor int8 quantization breaks this model; per-channel weights keep it accurate. On our synthetic test set it removed 97.8% of personal identifiers under ONNX Runtime 1.28.
• spm.model and tokenizer.json are the tokenizer of mDeBERTa-v3 base (commit a0484667b22365f84929a935b5e50a51f71f159d), by Microsoft, under the MIT licence, spm.model is unchanged; tokenizer.json was converted from it by the Hugging Face transformers library and adds GLiNER's two tokens, <<ENT>> and <<SEP>>.

Checksums (SHA-256)

File	Bytes	SHA-256
parakeet-tdt-0.6b-v3-pc8-2026-10-15/encoder.int8.onnx	881498555	c5cdff28d4ac8690b3e3f7780ac5089586d87a2800cebf0ed5d5a128e94f2b86
parakeet-tdt-0.6b-v3-pc8-2026-10-15/decoder.onnx	47233743	d593cdb0e571f5a457ec2219af9968cbf6b0e8198e8f7839b40a8754593bf68c
parakeet-tdt-0.6b-v3-pc8-2026-10-15/joiner.onnx	25286330	b9b0bcf88ac571902e69a6536223ed2d94885e981b85045410f1403d53121a63
parakeet-tdt-0.6b-v3-pc8-2026-10-15/tokens.txt	93939	d58544679ea4bc6ac563d1f545eb7d474bd6cfa467f0a6e2c1dc1c7d37e3c35d
parakeet-tdt-0.6b-v3-pc8-2026-10-15/silero_vad.onnx	643854	9e2449e1087496d8d4caba907f23e0bd3f78d91fa552479bb9c23ac09cbb1fd6
nemotron-3-diarization-q8-2026-09-23/Nemotron-3-Diarization.q8_0.gguf	107012128	08456d9e22cd9a323c0364d98375f3746d6e68507ebb705cd46438c534c7a3a1
gliner-multi-pii-v1-wo8-2026-10-24/model.onnx	365533791	b7569cce986a70955ba1b3da7b4c98e9cbef2c7e91b657ccf6498a7077110e8b
gliner-multi-pii-v1-wo8-2026-10-24/tokenizer.json	16034824	c5b8041501fcdee792b9b112dd592861f04391a24633a73fd9f05aaaf6e8eff1
gliner-multi-pii-v1-wo8-2026-10-24/spm.model	4305025	13c8d666d62a7bc4ac8f040aab68e942c861f93303156cc28f5c7e885d86d6e3
gliner-multi-pii-v1-wo8-2026-10-24/gliner_config.json	2312	5645ea576d26777a204be05a8cb97d9c2b15d04496c7712a6bb094c828ef1d9d
gliner-multi-pii-v1-pc8e-2026-10-23/model.onnx	354286656	92fdb25ad1bcc07476cb86d0e3aeafe38a549ad4f2664328daaab00e26237696
gliner-multi-pii-v1-pc8e-2026-10-23/tokenizer.json	16034824	c5b8041501fcdee792b9b112dd592861f04391a24633a73fd9f05aaaf6e8eff1
gliner-multi-pii-v1-pc8e-2026-10-23/spm.model	4305025	13c8d666d62a7bc4ac8f040aab68e942c861f93303156cc28f5c7e885d86d6e3
gliner-multi-pii-v1-pc8e-2026-10-23/gliner_config.json	2312	5645ea576d26777a204be05a8cb97d9c2b15d04496c7712a6bb094c828ef1d9d

The models are provided as is, with no warranty, under their own licences. A new version of a model is published as a new folder; existing folders are never changed.