---
language: [en, zh, fr, it, ko, pt, vi]
license: mit
pipeline_tag: automatic-speech-recognition
library_name: ggml
base_model: microsoft/VibeVoice-ASR-BitNet
tags: [ASR, quantization, cpu-inference, gguf, bitnet, multilingual, thox]
---

# ThoxSpeech ASR (BitNet)

**These are not new weights.** ThoxSpeech redistributes the quantized GGUF artifacts
from [microsoft/VibeVoice-ASR-BitNet](https://huggingface.co/microsoft/VibeVoice-ASR-BitNet)
**byte for byte**, under a THOX name, so the ThoxSpeech runtime and Space have a
stable artifact to pull.

THOX.ai did **not** train, fine-tune, re-quantize, distill, merge, or otherwise alter
these models. Any claim about their accuracy is a claim about Microsoft's work.

## Provenance and integrity

| ThoxSpeech filename | Upstream filename | Size | SHA-256 |
|---|---|---|---|
| `thoxspeech-vae-encoder-i8_s.gguf` | `vibeasr-vae-encoder-i8_s.gguf` | 703,080,064 B | `4941c82608c253ec066b5cc74d3dd11a5c8fef96cccbc5b87359ef0fe4338df6` |
| `thoxspeech-lm-i2_s-embed-q6_k.gguf` | `vibeasr-lm-i2_s-embed-q6_k.gguf` | 992,877,600 B | `fbe273d8dc2f2433bb25f849e19d77ea65aaa2188d12c20cee987ab6f321e002` |

Verify the identity claim yourself:

```bash
huggingface-cli download microsoft/VibeVoice-ASR-BitNet \
    vibeasr-vae-encoder-i8_s.gguf vibeasr-lm-i2_s-embed-q6_k.gguf --local-dir upstream
huggingface-cli download Thox-ai/ThoxSpeech-ASR-BitNet \
    thoxspeech-vae-encoder-i8_s.gguf thoxspeech-lm-i2_s-embed-q6_k.gguf --local-dir thox

sha256sum upstream/*.gguf thox/*.gguf   # digests must pair up across the two sets
```

`scripts/fetch-models.sh` performs this check automatically and refuses to continue
on a mismatch.

Digests recorded 2026-07-24 against `microsoft/VibeVoice-ASR-BitNet` @ `main`.

## Architecture

Two models, quantized differently on purpose — **heterogeneous quantization**:

| Stage | Component | Format | Size | Rationale |
|---|---|---|---|---|
| 1 | VAE acoustic + semantic tokenizer | **I8_S** (8-bit) | 0.65 GB | Audio features are precision-sensitive. Quantize here too hard and the signal is gone before the decoder sees it. |
| 2 | LM decoder (Qwen2.5-1.5B) | **I2_S** ternary + Q6_K embeddings | 0.92 GB | Autoregressive decoding is memory-bandwidth-bound; ternary weights turn multiplies into adds and cut the working set 3.6x. Embeddings stay at Q6_K because the token table does not tolerate ternary. |

The decoder is Qwen2.5-**1.5B**, substituted upstream for the 7B used in the
full-size VibeVoice-ASR. That costs 1–4% absolute WER and takes the total from
4.62 GB to 1.58 GB.

## Intended use

Speech-to-text on CPU-only hardware — edge devices, laptops, free-tier containers —
where a GPU is unavailable and a network round-trip is undesirable. In ThoxOS this
backs [ThoxRoute](https://github.com/ttracx/thoxroute)'s speech-recognition route,
local-first.

**Languages:** English, Chinese, French, Italian, Korean, Portuguese, Vietnamese.

## Limitations

Inherited from the upstream model; none of these are ThoxSpeech-specific:

- **Quantization costs accuracy.** 1–4% absolute WER above the FP16 7B model. On
  hard meeting audio the gap is much wider — AISHELL4 goes 19.83 → 27.45 WER.
- **Far-field and overlapping speech degrade badly.** AliMeeting sits at 40.58 WER.
  This is not a model for noisy multi-speaker rooms.
- **English and Chinese are strongest.** Portuguese (24.87) and Vietnamese (22.38)
  are materially worse. The seven listed languages are the only supported ones;
  anything else is undefined behaviour, not graceful degradation.
- **Whisper and Parakeet still win on clean read speech.** LibriSpeech-clean: 2.41
  here vs. 1.49 for Parakeet. ThoxSpeech's advantage is speed per CPU watt, not
  peak accuracy.
- **No diarization, no timestamps, no punctuation guarantees.**
- **24 kHz mono input.** Other rates degrade accuracy silently rather than erroring,
  which is why the ThoxSpeech server force-normalizes every input through ffmpeg.

## Bias and risk

Trained on data THOX has not audited. Expect the usual ASR failure modes — higher
error rates for accented speech, dialects, code-switching, children's voices, and
atypical speech — distributed unevenly across speaker populations. Do not use
transcripts as evidence in any consequential decision about a person without human
review. See Microsoft's model card for whatever training-data detail exists upstream.

## License

**MIT**, inherited from `microsoft/VibeVoice-ASR-BitNet`. Copyright (c) Microsoft
Corporation.

The decoder derives from **Qwen2.5-1.5B** (Apache-2.0, © Alibaba Cloud). Apache-2.0
§4(b) modification notice: the weights were quantized to I2_S ternary with Q6_K
embeddings and converted to GGUF, upstream of THOX.

Redistribution must retain both notices. See
[`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md) and [`NOTICE`](NOTICE).

THOX.ai LLC is not affiliated with, sponsored by, or endorsed by Microsoft
Corporation or Alibaba Cloud.

## Citation

Cite the original authors, not THOX:

```bibtex
@article{xu2025vibeasrbitnet,
    title={VibeVoice-ASR-BitNet Technical Report},
    author={Xu, Songchen and Song, Ting and Huang, Shaohan and Peng, Zhiliang and
            Xia, Yan and Tu, Yujie and Huang, Xin and Yu, Jianwei and Dong, Li and
            Wei, Furu},
    journal={arXiv preprint arXiv:2607.21075},
    year={2025}
}
```