ThoxSpeech-ASR-BitNet

Your AI. Your Data. Your Rules.

Verbatim redistribution of microsoft/VibeVoice-ASR-BitNet — CPU-only multilingual ASR (NOT retrained or re-quantized by THOX).

What this is

• This is a byte-for-byte mirror of microsoft/VibeVoice-ASR-BitNet (MIT, © Microsoft) with a SHA-256 provenance table. The decoder derives from Qwen2.5-1.5B (Apache-2.0, © Alibaba).
• Total 1.58 GB (I8_S acoustic tokenizer 0.65 GB + I2_S/Q6_K LM 0.92 GB). CPU-only.
• Cite Microsoft, not THOX. Accuracy degrades on far-field / overlapping speech (internal note: WER ~2.41 on LibriSpeech-clean).

Architecture (from config)

Field	Value
Architecture	2-stage: I8_S VAE tokenizer + I2_S/Q6_K Qwen2.5-1.5B LM decoder

Intended use

Multilingual automatic speech recognition on CPU. Research / on-device transcription.

Usage

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

Links

• Base model: microsoft/VibeVoice-ASR-BitNet
• Docs: https://docs.thox.ai

Files

File	Size	SHA-256
thoxspeech-lm-i2_s-embed-q6_k.gguf	992.88 MB (992,877,600 B)	fbe273d8dc2f2433bb25f849e19d77ea65aaa2188d12c20cee987ab6f321e002
thoxspeech-vae-encoder-i8_s.gguf	703.08 MB (703,080,064 B)	4941c82608c253ec066b5cc74d3dd11a5c8fef96cccbc5b87359ef0fe4338df6

Use with ThoxOS

Download the file once and run it fully offline on a ThoxOS host (or any Linux/macOS/Windows machine):

pip install -U "huggingface_hub[cli]"
hf download Thox-ai/ThoxSpeech-ASR-BitNet thoxspeech-lm-i2_s-embed-q6_k.gguf --local-dir ./ThoxSpeech-ASR-BitNet
llama-cli -m ./ThoxSpeech-ASR-BitNet/thoxspeech-lm-i2_s-embed-q6_k.gguf -p "Hello" -n 128

Ollama / LM Studio: not supported for this file (ternary TQ* or non-standard GGUF types need a recent llama.cpp build).

Use with MeshStack

Serve the model as an OpenAI-compatible endpoint on any MeshStack node, then register that endpoint with your MeshStack coordinator (protocol openai, base URL http://<node-ip>:8080/v1; or protocol ollama at http://<node-ip>:11434 if you serve it with Ollama).

llama-server -m ./ThoxSpeech-ASR-BitNet/thoxspeech-lm-i2_s-embed-q6_k.gguf --host 0.0.0.0 --port 8080 -c 4096
curl http://<node-ip>:8080/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Hello"}]}'

Device fit (by file size only)

• thoxspeech-lm-i2_s-embed-q6_k.gguf (992.88 MB): Medium (<2.5 GB): devices with ≥4 GB RAM — recent phones, laptops, Apple Silicon Macs, MeshStack edge hosts.

RAM needed ≈ file size + context (KV cache) + runtime overhead. These tiers are derived from file size alone; no throughput, TOPS or battery figures are claimed here.

Licence and attribution

Mirror/conversion of microsoft/VibeVoice-ASR-BitNet (MIT, © Microsoft Corporation). The MIT notice carries forward; THOX claims no authorship of the weights.

Out of scope

• Medical, legal, financial, safety-critical or other high-stakes decisions without human review.
• Unattended control of physical devices or actuators without a confirmation step.
• Any use prohibited by the upstream licence or acceptable-use policy.

Limitations and bias

• Small and quantized models make factual mistakes and can produce confident but wrong output; verify important answers.
• Behaviour and bias reflect the training data, which is mostly English web/instruction text; quality in other languages is lower unless stated above.
• Quantization trades accuracy for size; lower-bit files degrade more.

---

THOX.ai LLC — Your AI. Your Data. Your Rules.™ · thox.ai · Contact: tommy@thox.ai