Modelpublic

geantendormi/MOSS-Transcribe-Diarize-Q4_K_M.gguf

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes43downloads

Model Card

MOSS-Transcribe-Diarize-GGUF (Q4KM Hybrid)

GGUF conversions of OpenMOSS/MOSS-Transcribe-Diarize.

Joint ASR + Speaker Diarization + Timestamps in a single 0.9B Speech-LLM model. Produces timestamped, speaker-labelled transcripts in one pass.

📦 Model Files & Quantization Architecture

File	Size	Description
`MOSS-Transcribe-Diarize-Q4_K_M.gguf`	1.02 GB	SOTA Hybrid Quantization: Whisper 24L Encoder (F16) + VQAdaptor (F16) + Qwen3 28L LM Decoder (Q4KM)

🛠️ Hybrid Quantization Specifications

—Acoustic Encoder (Whisper-Medium 24L): Preserved at F16 (16-bit) precision to maintain zero audio signal SNR degradation.
—VQAdaptor Feature Projection: Preserved at F16 (16-bit) precision.
—LM Decoder (Qwen3-0.6B 28L): Quantized to Q4_K_M (4-bit) super-block representation for fast Token generation.
—Header Metadata: MODEL_ARCH.QWEN3, key_length = 128, QK-Norm (attn_q_norm & attn_k_norm) fully enabled.

🚀 Quickstart with llama.cpp

Run Command

bash

llama-cli -m MOSS-Transcribe-Diarize-Q4_K_M.gguf -p "请将音频转写为文本：" -n 32 -t 8

Output Format Example

text

[00:00:00.250 --> 00:00:01.620] [S01] Can we get a table for two?
[00:00:01.710 --> 00:00:07.010] [S02] Sure. One table is just about to open up, just a few minutes.
[00:00:07.140 --> 00:00:07.640] [S01] Thank you.

📄 License

Apache-2.0 (same as the base model).