ThoxSpeech-ASR-BitNet

CPU Multilingual Speech Recognition (Redistribution) — a verbatim redistribution of a third-party multilingual automatic speech recognition model, mirrored here for CPU-only on-device transcription in the THOX stack

Multilingual ASR, CPU-only · 1.58 GB total (I8_S tokenizer 0.65 GB + I2_S / Q6_K decoder 0.92 GB) · Verbatim upstream redistribution

Your AI. Your Data. Your Rules.™

Status: VERBATIM REDISTRIBUTION. THOX did not retrain or re-quantize this model; it is mirrored here for CPU-only on-device transcription in the THOX stack, under the upstream MIT license, with credit to the upstream authors (see License).

Summary

ThoxSpeech-ASR-BitNet is a verbatim redistribution of a third-party multilingual automatic speech recognition model, mirrored here for CPU-only on-device transcription in the THOX stack. THOX did not retrain or re-quantize it; credit for the model belongs upstream (see License). Total size 1.58 GB: an I8_S acoustic tokenizer (0.65 GB) plus an I2_S / Q6_K language-model decoder (0.92 GB), in a two-stage architecture.

Highlights

• ThoxSpeech-ASR-BitNet is a verbatim redistribution of a third-party multilingual automatic speech recognition model, mirrored here for CPU-only on-device transcription in the THOX stack.
• bash hf download Thox-ai/ThoxSpeech-ASR-BitNet --local-dir ./ThoxSpeech-ASR-BitNet # run with the VibeVoice / BitNet ggml ASR runtime (CPU) Ollama and LM Studio do not support these files; the ternary and non-standard GGUF types need the dedicated ggml ASR runtime.

Specifications

Files

Run it

hf download Thox-ai/ThoxSpeech-ASR-BitNet --local-dir ./ThoxSpeech-ASR-BitNet
# run with the VibeVoice / BitNet ggml ASR runtime (CPU)

Ollama and LM Studio do not support these files; the ternary and non-standard GGUF types need the dedicated ggml ASR runtime.

Platforms

Intended use

• Multilingual automatic speech recognition on CPU: research and on-device transcription.
• Out of scope: medical, legal, financial, safety-critical, or other high-stakes decisions without human review; unattended control of physical devices without a confirmation step.

Training data and procedure

THOX fine-tune of microsoft/VibeVoice-ASR-BitNet. The corpus and procedure are described in this card as published at release (see Summary and Evaluation); no training beyond what this card describes is claimed for this repository.

Evaluation

Word error rate around 2.41 on LibriSpeech-clean in THOX's internal testing. Accuracy degrades on far-field and overlapping speech. No THOX GATE49 result applies (speech model).

Limitations

• Transcription errors are expected on noisy, far-field, or overlapping speech; verify important transcripts.
• THOX makes no quality claim beyond the measurement above; the model is redistributed as published upstream.
• Behaviour and bias reflect the upstream training data.

License

MIT. This repo is a mirror of an upstream MIT-licensed model (© Microsoft Corporation); the MIT notice carries forward and THOX claims no authorship of the weights. See the license text for terms.

THOX.ai — Your AI. Your Data. Your Rules.™

Provenance

Published by THOX.ai as Thox-ai/ThoxSpeech-ASR-BitNet. Upstream base: microsoft/VibeVoice-ASR-BitNet. This card was conformed to THOX Model Card Standard v1.0 on 2026-10-31 (frontmatter, tags, header, files, and section structure); the technical content is as published at release.

Citation

@misc{thoxai2026thoxspeechasrbitnet,
      title={ThoxSpeech-ASR-BitNet},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxSpeech-ASR-BitNet}
}