schift-io/vibevoice7b-schift-quanted-mixed

VibeVoice-ASR-Streaming-7B — Mixed-Precision Quantized (6GB, self-contained)

Experimental artifact, but self-contained: this repo alone is enough to run real speech-to-text with speaker diarization — no separate download of microsoft/VibeVoice-ASR-Streaming-7B is needed. It bundles:

• vibevoice7b-schift-quanted-mixed.gguf — the Qwen2 text decoder + lm_head, mixed-precision quantized to 6GB (was ~14GB at bf16).
• audio_components.safetensors (2.1GB, bf16) — the acoustic tokenizer, semantic tokenizer, and both connectors, unmodified from the original checkpoint.
• config.json + tokenizer files — everything load_full_model.py needs to reconstruct the full VibeVoiceASRForConditionalGeneration and call model.streaming_generate(...) directly.

Total download: ~8.4GB, vs. ~17GB for the original bf16 checkpoint.

What this is

• The decoder (language_model + lm_head) was extracted from the official 7B checkpoint and verified to match a vanilla Qwen2ForCausalLM exactly (0 missing / 0 unexpected keys).
• Quantized with llama.cpp, mixed per-tensor-kind precision guided by a Korean imatrix calibration built from the model's own JSON transcription schema:
  • ffn_gate / ffn_up → Q8_0
  • attn_k / attn_q → Q4_K
  • attn_v / attn_output / ffn_down → IQ3_XXS
  • token embedding / output → Q8_0
• Result: 6.02 GB decoder (vs. ~14GB at bf16), perplexity within noise of full-precision f16 on the calibration set (318.3 vs. 318.6).
• Audio components (tokenizers + connectors) are kept at bf16, unquantized — they're small (2.1GB) and this repo doesn't touch them beyond re-packaging.

Real-audio validation (not just perplexity)

Run end-to-end through model.streaming_generate() on 20 real Korean broadcast-panel-discussion utterances (AIHub, multi-speaker, continuous conversational context, character-level CER with speaker-label tokens stripped):

	CER
Original bf16 decoder	19.39%
This GGUF (6GB, mixed precision)	19.27%

Essentially identical — within sample noise (n=20). Quantization did not measurably hurt recognition quality on this test.

Speaker diarization validation

The CER numbers above strip Speaker N: labels before scoring — they measure content transcription only, not diarization. Diarization was measured separately with cpCER (concatenated minimum-permutation CER, the CHiME6/NOTSOFAR standard): text is grouped by speaker label (hypothesis vs. ground-truth), and the hypothesis-speaker→ground-truth-speaker mapping is searched over all permutations to find the one minimizing total CER (label IDs are arbitrary, so this permutation-invariant matching is required).

	cpCER (diarization + transcription combined)	Speaker-count mismatches
Original bf16 decoder	37.14%	10/20
This GGUF (6GB, mixed precision)	34.41%	10/20

Quantization does not measurably hurt diarization either — GGUF and original are on par (sample-level variance is large but there's no systematic direction).

⚠️ The base VibeVoice-7B checkpoint's own diarization is imperfect, independent of this quantization: on 10 of 20 samples it merges or splits the true speaker count (e.g. reports 2 speakers when there were 3). Samples where the speaker count is correct score 5-20% cpCER; samples where it's wrong jump to 28-137%. This is a base-model characteristic this repo inherits unchanged, not something introduced by quantization.

Usage

pip install torch transformers safetensors gguf
pip install git+https://github.com/microsoft/VibeVoice.git   # model code (not weights)

from huggingface_hub import snapshot_download
repo_dir = snapshot_download("schift-io/vibevoice7b-schift-quanted-mixed")

from load_full_model import load_full_model   # ships in this repo
model, processor = load_full_model(repo_dir)

# model.streaming_generate(...) — see the VibeVoice repo's
# demo/vibevoice_asr_streaming_inference_from_file.py for the full call pattern

load_full_model.py builds the architecture from config.json, loads the audio components from audio_components.safetensors, dequantizes the GGUF decoder in pure Python (gguf_to_qwen2model.py, via the gguf PyPI package — no llama.cpp binary needed), and wires it all into one ready-to-run model. No network access to the original checkpoint is required.

What this is not

• Not officially supported by Microsoft. Independent experiment, shared as-is.
• Not a speaker-diarization/head-split architecture — that work is separate, unpublished, and still experimental (untrained/unvalidated).
• No further quantization variants (Q4_K_M, IQ2_XXS, etc. were tried and were worse; this mixed config was the best found).

License

MIT, inherited from the base model (microsoft/VibeVoice-ASR-Streaming-7B). See LICENSE.