AnalogVector/voxdicta-models

Voxdicta model artifacts

Public mirror of the model files consumed by Voxdicta. Files retain their upstream filenames. This repository combines components under different terms; no single license applies to every file.

Files, sources, and terms

• ggml-large-v3-turbo-q8_0.bin — Whisper Large v3 Turbo Q8_0 artifact. Source: ggerganov/whisper.cpp; the upstream model card identifies the model as MIT-licensed.
• canary-1b-v2-Q8_0.gguf — Q8_0 GGUF conversion of NVIDIA Canary 1B v2 (model_weights.ckpt of the .nemo archive): the FastConformer encoder and the Transformer decoder, with convolutions, norms and biases kept in F32 and the linear matrices quantized to 8 bits. Converted and quantized by this repository's owner (changes made to the original); Canary is © NVIDIA, CC-BY-4.0.
• canary-1b-v2-timestamps-Q8_0.gguf — Q8_0 GGUF conversion of the CTC timestamps model that ships inside the same .nemo archive (timestamps_asr_model_weights.ckpt), used to align words. Converted and quantized by this repository's owner (changes made to the original); © NVIDIA, CC-BY-4.0.
• timestamps-model.int8.onnx, timestamps-vocab.txt, timestamps-tokenizer.model — CTC timestamp assets (the ONNX model aligns words on the CPU; the vocabulary and tokenizer serve both aligners) attributed to ysdede/nemo-canary-1b-v2-onnx-timestamped, CC-BY-4.0.
• Qwen3-ASR-1.7B-Q8_0.gguf — Q8_0 GGUF conversion of Qwen3-ASR-1.7B (© Alibaba Cloud, Qwen team), licensed under Apache License 2.0; the licence text is in LICENSES/Qwen3-ASR-Apache-2.0.txt. Converted and quantized to 8 bits by sakasegawa (changes made to the original), and mirrored here unchanged (SHA-256 5f219b78a1d9c3b9e97da27708b36f8a0bc1bfc1650b541c0a6dbaf87c9a62d0).
• silero_vad.int8.onnx — Silero VAD artifact from the sherpa-onnx ASR model release; see the upstream release for its terms.
• silero-vad-309K-v6.2-F32.gguf — Silero VAD v6.2 (16 kHz, F32) in the layout of speech.cpp, © Silero Team, licensed under the MIT License (LICENSES/Silero-VAD-MIT.txt). Converted by sakasegawa (changes made to the original) and mirrored here unchanged (SHA-256 2fc312514e5fc4fa4ffd02b25c76a4853fa4130e8fb9777d01502183b01238cb). The Windows application's Canary uses it to find the regions that contain speech.
• qwen3-forced-aligner-0.6b-q8_0.gguf — Q8_0 GGUF of Qwen3-ForcedAligner-0.6B (© Alibaba Cloud, Qwen team), licensed under Apache License 2.0 (the same text as LICENSES/Qwen3-ASR-Apache-2.0.txt). Converted and quantized to 8 bits by cstr (changes made to the original) and mirrored here unchanged (SHA-256 539df5dd0fe1721e378ac13bfac9a26b1260dafb62d892c518c1f21244762636). The Windows application uses it for word timestamps with every recognition model.
• Nemotron-3-Diarization-99M-Q8_0.gguf — Q8_0 GGUF of NVIDIA Nemotron 3 Diarization (© NVIDIA Corporation) in the layout of speech.cpp, licensed under the OpenMDW License Agreement, version 1.1 (LICENSES/OpenMDW-1.1.txt). Converted and quantized to 8 bits by sakasegawa (changes made to the original), and mirrored here unchanged (SHA-256 336df6350dce2ea5ff06aa6e2b249d331a859ced13305fbb45c44bdb3c0abe20). The Windows application runs it.
• Nemotron-3-Diarization.q8_0.gguf — Q8_0 GGUF of the same model as NVIDIA publishes it (NeMo-Speech.cpp layout), licensed under the OpenMDW License Agreement, version 1.1 (LICENSES/OpenMDW-1.1.txt). The Android application runs it.

This is an independent mirror, not an official NVIDIA, OpenAI, or sherpa-onnx repository. Preserve the applicable upstream attributions and licenses when redistributing any file.