Dubedo/VibeVoice-ASR-HF-INT8

VibeVoice-ASR-HF — Selective INT8 8-bit Quantization

Selectively quantized version of microsoft/VibeVoice-ASR-HF for low-VRAM deployment.

Only the Qwen2.5-7B LLM backbone is quantized. The acoustic tokenizer encoder, semantic tokenizer encoder, projection layers, and lm_head remain in full BF16 precision — preserving diarization accuracy and transcription quality.

Key details

Base model	microsoft/VibeVoice-ASR-HF
Quantization	INT8 8-bit (bitsandbytes)
Modules quantized	language_model.model.layers.* only
Modules in BF16	acoustic_tokenizer_encoder, semantic_tokenizer_encoder, acoustic_projection, semantic_projection, lm_head
Model size	~9 GB (down from 17.3 GB)
VRAM usage	~10–11 GB
Transformers	>= 5.3.0
bitsandbytes	>= 0.48.1

Why selective quantization?

Naive 8-bit quantization of the entire model destroys diarization (all speakers collapse to SPEAKER_00) and degrades transcription quality significantly. The acoustic and semantic tokenizer encoders process raw audio signals where small numerical errors propagate catastrophically through the convolutional stages. The LLM backbone (Qwen2.5-7B) handles quantization gracefully since its weights follow a normal distribution well-suited for INT8.

Usage

import torch
from transformers import AutoProcessor, VibeVoiceAsrForConditionalGeneration

model_id = "Dubedo/VibeVoice-ASR-HF-INT8"

processor = AutoProcessor.from_pretrained(model_id)
model = VibeVoiceAsrForConditionalGeneration.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

inputs = processor.apply_transcription_request(
    audio="path/to/audio.wav",
    prompt="optional hotwords here",
).to(model.device, model.dtype)

output_ids = model.generate(**inputs)
generated_ids = output_ids[:, inputs["input_ids"].shape[1]:]

# Structured output with speaker, timestamps, text
result = processor.decode(generated_ids, return_format="parsed")[0]

Quantization method

Quantized using BitsAndBytesConfig with llm_int8_skip_modules to protect audio-critical components:

BitsAndBytesConfig(
    load_in_8bit=True,
    llm_int8_skip_modules=[
        "acoustic_tokenizer_encoder",
        "semantic_tokenizer_encoder",
        "acoustic_projection",
        "semantic_projection",
        "lm_head",
    ],
)

Acknowledgments

Based on the selective quantization approach documented by FabioSarracino/VibeVoice-Large-Q8 and Enemyx-net/VibeVoice-ComfyUI, adapted for the HF-native ASR architecture in transformers 5.3.0.