VibeVoice-AcousticTokenizer

VibeVoice Acoustic Tokenizer

VibeVoice is a novel framework designed for generating expressive, long-form, multi-speaker conversational audio, such as podcasts, from text. It addresses significant challenges in traditional Text-to-Speech (TTS) systems, particularly in scalability, speaker consistency, and natural turn-taking.

A core innovation of VibeVoice is its use of continuous speech tokenizers (Acoustic and Semantic) operating at an ultra-low frame rate of 7.5 Hz. These tokenizers efficiently preserve audio fidelity while significantly boosting computational efficiency for processing long sequences. VibeVoice employs a next-token diffusion framework, leveraging a Large Language Model (LLM) to understand textual context and dialogue flow, and a diffusion head to generate high-fidelity acoustic details.

The speech tokenizer is a key component for both VibeVoice TTS and ASR.

➡️ Technical Report: VibeVoice Technical Report

➡️ Project Page: microsoft/VibeVoice

<p align="left"> <img src="figs/tokenizer_comparison.png" alt="Tokenizer Comparison" height="250px"> </p>

Models

Usage

Setup

Until the VibeVoice acoustic tokenizer is part of an official Transformers release, it can be used by installing from the source code:

pip install git+https://github.com/huggingface/transformers.git

Example

<details> <summary>Encoding and decoding</summary>

import torch
from scipy.io import wavfile

from transformers import AutoFeatureExtractor, VibeVoiceAcousticTokenizerModel
from transformers.audio_utils import load_audio_librosa


model_id = "microsoft/VibeVoice-AcousticTokenizer"

# load model
feature_extractor = AutoFeatureExtractor.from_pretrained(model_id)
model = VibeVoiceAcousticTokenizerModel.from_pretrained(model_id, device_map="auto")
print("Model loaded on device:", model.device)
print("Model dtype:", model.dtype)

# load audio
audio = load_audio_librosa(
    "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Alice_woman.wav",
    sampling_rate=feature_extractor.sampling_rate,
)

# preprocess audio
inputs = feature_extractor(
    audio,
    sampling_rate=feature_extractor.sampling_rate,
    pad_to_multiple_of=3200,
).to(model.device, model.dtype)
print("Input audio shape:", inputs.input_values.shape)
# Input audio shape: torch.Size([1, 1, 224000])

with torch.no_grad():
    # set VAE sampling to False for deterministic output
    encoded_outputs = model.encode(inputs.input_values, sample=False)
    print("Latent shape:", encoded_outputs.latents.shape)
    # Latent shape: torch.Size([1, 70, 64])

    decoded_outputs = model.decode(**encoded_outputs)
    print("Reconstructed audio shape:", decoded_outputs.audio.shape)
    # Reconstructed audio shape: torch.Size([1, 1, 224000])

# Save audio
output_fp = "vibevoice_acoustic_tokenizer_reconstructed.wav"
wavfile.write(output_fp, feature_extractor.sampling_rate, decoded_outputs.audio.squeeze().float().cpu().numpy())
print(f"Reconstructed audio saved to : {output_fp}")

</details>

Original audio <audio controls> <source src="https://hf.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Alice_woman.wav" type="audio/wav"> </audio>

Encoded/decoded audio <audio controls> <source src="https://hf.co/datasets/bezzam/vibevoice_samples/resolve/main/example_output/vibevoice_acoustic_tokenizer_reconstructed.wav" type="audio/wav"> </audio>

<details> <summary>Streaming</summary>

For streaming ASR or TTS, where cached states need to be tracked, the use_cache parameter can be used when encoding or decoding audio:

import torch
from scipy.io import wavfile

from transformers import AutoFeatureExtractor, VibeVoiceAcousticTokenizerModel
from transformers.audio_utils import load_audio_librosa


model_id = "microsoft/VibeVoice-AcousticTokenizer"


# load model
feature_extractor = AutoFeatureExtractor.from_pretrained(model_id)
model = VibeVoiceAcousticTokenizerModel.from_pretrained(model_id, device_map="auto")
print("Model loaded on device:", model.device)
print("Model dtype:", model.dtype)

# load audio
audio = load_audio_librosa(
    "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Alice_woman.wav",
    sampling_rate=feature_extractor.sampling_rate,
)

# preprocess audio
inputs = feature_extractor(
    audio,
    sampling_rate=feature_extractor.sampling_rate,
    pad_to_multiple_of=3200,
).to(model.device, model.dtype)
print("Input audio shape:", inputs.input_values.shape)
# Input audio shape: torch.Size([1, 1, 224000])

# chache will be initialized after a first pass
encoder_cache = None
decoder_cache = None
with torch.no_grad():
    # set VAE sampling to False for deterministic output
    encoded_outputs = model.encode(inputs.input_values, sample=False, padding_cache=encoder_cache, use_cache=True)
    print("Latent shape:", encoded_outputs.latents.shape)
    # Latent shape: torch.Size([1, 70, 64])

    decoded_outputs = model.decode(encoded_outputs.latents, padding_cache=decoder_cache, use_cache=True)
    print("Reconstructed audio shape:", decoded_outputs.audio.shape)
    # Reconstructed audio shape: torch.Size([1, 1, 224000])

    # `padding_cache` can be extracted from the outputs for subsequent passes
    encoder_cache = encoded_outputs.padding_cache
    print("Number of cached encoder layers:", len(encoder_cache.per_layer_in_channels))
    # Number of cached encoder layers: 34
    decoder_cache = decoded_outputs.padding_cache
    print("Number of cached decoder layers:", len(decoder_cache.per_layer_in_channels))
    # Number of cached decoder layers: 34

# Save audio
output_fp = "vibevoice_acoustic_tokenizer_reconstructed.wav"
wavfile.write(output_fp, feature_extractor.sampling_rate, decoded_outputs.audio.squeeze().float().cpu().numpy())
print(f"Reconstructed audio saved to : {output_fp}")

</details>