VibeVoice Technical Report
Paper • 2508.19205 • Published • 177
VibeVoice-7B-hf (Transformers-compatible version)
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 model can synthesize speech up to 45 minutes long with up to 4 distinct speakers, surpassing the typical 1-2 speaker limits of many prior models.
➡️ Technical Report: VibeVoice Technical Report
➡️ Project Page: microsoft/VibeVoice
This model was contributed by Eric Bezzam.
Usage
Setup
Until VibeVoice is part of a Transformers release, you can install it from source:
pip install git+https://github.com/huggingface/transformers.git
A noise scheduler is needed as audio generation relies on a diffusion process. By default, the model will create a noise scheduler with diffusers internally.
pip install diffusers
pip install soundfile # for saving audio
Loading the model
from transformers import AutoProcessor, AutoModelForTextToWaveform
model_id = "microsoft/VibeVoice-7B-hf"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTextToWaveform.from_pretrained(model_id)
Text-to-speech (TTS)
import os
from transformers import AutoProcessor, AutoModelForTextToWaveform
model_id = "microsoft/VibeVoice-7B-hf"
text = "Hello, nice to meet you. How are you?"
# Load model
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTextToWaveform.from_pretrained(model_id, device_map="auto")
# Prepare input
conversation = [{"role": "0", "content": [{"type": "text", "text": text}]}]
inputs = processor.apply_chat_template(
conversation, return_dict=True, tokenize=True, add_generation_prompt=True,
).to(model.device, model.dtype)
# Generate!
audio = model.generate(**inputs)
# Save to file
file_name = f"{os.path.basename(model_id)}_tts.wav"
processor.save_audio(audio, file_name)
print(f"Saved output to {file_name}")
TTS voice cloning
A voice can be cloned by providing a reference audio alongside the text within the chat template dictionary.
import os
from transformers import AutoProcessor, AutoModelForTextToWaveform, set_seed
model_id = "microsoft/VibeVoice-7B-hf"
text = "Hello, nice to meet you. How are you?"
set_seed(42) # for deterministic results
# Load model
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTextToWaveform.from_pretrained(model_id, device_map="auto")
sampling_rate = processor.feature_extractor.sampling_rate
# Prepare input
conversation = [
{
"role": "0",
"content": [
{"type": "text", "text": text},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Alice_woman.wav",
},
],
}
]
inputs = processor.apply_chat_template(
conversation, return_dict=True, tokenize=True, add_generation_prompt=True,
).to(model.device, model.dtype)
# Generate!
audio = model.generate(**inputs)
# Save to file
fn = f"{os.path.basename(model_id)}_tts_clone.wav"
processor.save_audio(audio, fn)
print(f"Saved output to {fn}")
Generating a podcast from a script
Below is an example to generate a conversation between two speakers, whose voices are cloned by providing a reference audio for each unique role ID in the chat template.
The example below also uses the monitor_progress option to track the generation progress.
import os
import time
from transformers import AutoProcessor, AutoModelForTextToWaveform
model_id = "microsoft/VibeVoice-7B-hf"
max_new_tokens = 400 # `None` to ensure full generation
# create conversation with an audio for the first time a speaker appears to clone that particular voice
conversation = [
{
"role": "0",
"content": [
{
"type": "text",
"text": "Hello everyone, and welcome to the VibeVoice podcast. I'm your host, Linda, and today we're getting into one of the biggest debates in all of sports: who's the greatest basketball player of all time? I'm so excited to have Thomas here to talk about it with me.",
},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Alice_woman.wav",
},
],
},
{
"role": "1",
"content": [
{
"type": "text",
"text": "Thanks so much for having me, Linda. You're absolutely right—this question always brings out some seriously strong feelings.",
},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Frank_man.wav",
},
],
},
{
"role": "0",
"content": [
{
"type": "text",
"text": "Okay, so let's get right into it. For me, it has to be Michael Jordan. Six trips to the Finals, six championships. That kind of perfection is just incredible.",
},
],
},
{
"role": "1",
"content": [
{
"type": "text",
"text": "Oh man, the first thing that always pops into my head is that shot against the Cleveland Cavaliers back in '89. Jordan just rises, hangs in the air forever, and just sinks it",
},
],
},
]
# Load model
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTextToWaveform.from_pretrained(model_id, device_map="auto")
# prepare inputs
inputs = processor.apply_chat_template(
conversation, return_dict=True, tokenize=True, add_generation_prompt=True,
).to(model.device, model.dtype)
# Generate audio with a progress bar to track generation
model.generation_config.max_new_tokens = max_new_tokens
start_time = time.time()
audio = model.generate(**inputs, monitor_progress=True)
generation_time = time.time() - start_time
print(f"Generation time: {generation_time:.2f} seconds")
# Save audio
fn = f"{os.path.basename(model_id)}_script.wav"
processor.save_audio(audio, fn)
print(f"Saved output to {fn}")
Batched inference
For batch processing, a list of conversations can be passed to processor.apply_chat_template:
import os
import time
from transformers import AutoProcessor, AutoModelForTextToWaveform
model_id = "microsoft/VibeVoice-7B-hf"
max_new_tokens = 400 # `None` to ensure full generation
conversation = [
[
{
"role": "0",
"content": [
{
"type": "text",
"text": "Hello everyone, and welcome to the VibeVoice podcast. I'm your host, Linda, and today we're getting into one of the biggest debates in all of sports: who's the greatest basketball player of all time? I'm so excited to have Thomas here to talk about it with me.",
},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Alice_woman.wav",
},
],
},
{
"role": "1",
"content": [
{
"type": "text",
"text": "Thanks so much for having me, Linda.",
},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Frank_man.wav",
},
],
},
],
[
{
"role": "0",
"content": [
{
"type": "text",
"text": "Hello and welcome to Planet in Peril. I'm your host, Alice. We're here today to discuss a really sobering new report that looks back at the last ten years of climate change. I'm joined by our expert panel. Welcome Carter, Frank, and Maya.",
},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Alice_woman.wav",
},
],
},
{
"role": "1",
"content": [
{"type": "text", "text": "Hi Alice, it's great to be here. I'm Carter."},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Carter_man.wav",
},
],
},
{
"role": "2",
"content": [
{"type": "text", "text": "Hello, uh, I'm Frank. Good to be on."},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Frank_man.wav",
},
],
},
{
"role": "3",
"content": [
{"type": "text", "text": "And I'm Maya. Thanks for having me."},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Maya_woman.wav",
},
],
},
],
]
# Load model
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTextToWaveform.from_pretrained(model_id, device_map="auto")
# prepare inputs
inputs = processor.apply_chat_template(
conversation, return_dict=True, tokenize=True, add_generation_prompt=True,
).to(model.device, model.dtype)
# Generate audio with a progress bar to track generation
model.generation_config.max_new_tokens = max_new_tokens
start_time = time.time()
audio = model.generate(**inputs, monitor_progress=True)
generation_time = time.time() - start_time
print(f"Generation time: {generation_time:.2f} seconds")
# Save audio
output_dir = f"{os.path.basename(model_id)}_batch"
processor.save_audio(audio, output_dir)
print(f"Saved output to {output_dir}")
Pipeline usage
VibeVoice can also be loaded as a pipeline. We also show below how the diffusion parameters can be adjusted.
import os
import soundfile as sf
from transformers import pipeline
model_id = "microsoft/VibeVoice-7B-hf"
text = "Hello, nice to meet you. How are you?"
pipe = pipeline("text-to-speech", model=model_id)
# Generate!
conversation = [
{
"role": "0",
"content": [
{"type": "text", "text": text},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/voices/en-Alice_woman.wav",
},
],
}
]
# optional kwargs for generation
generate_kwargs = {"guidance_scale": 1.3, "num_diffusion_steps": 10}
output = pipe(conversation, generate_kwargs=generate_kwargs)
# Save to file
fn = f"{os.path.basename(model_id)}_pipeline.wav"
sf.write(fn, output["audio"], output["sampling_rate"])
print(f"Saved output to {fn}")
Training
VibeVoice can be trained with the loss outputted by the model.
from transformers import AutoProcessor, AutoModelForTextToWaveform
model_id = "microsoft/VibeVoice-7B-hf"
# Load model and processor
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTextToWaveform.from_pretrained(
model_id,
diffusion_loss_weight=0.75, # by default, equal weighting (0.5) of language modeling loss (CE) and diffusion loss is applied
device_map="auto"
)
model.train()
# Prepare batch of 2
conversation = [
[
{
"role": "0",
"content": [
{
"type": "text",
"text": "VibeVoice is this novel framework designed for generating expressive, long-form, multi-speaker, conversational audio.",
},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/realtime_model/vibevoice_tts_german.wav",
},
],
}
],
# NOTE: multiple speakers not supported yet
[
{
"role": "0",
"content": [
{
"type": "text",
"text": "Hello everyone and welcome to the VibeVoice podcast. I'm your host, Alex, and today we're getting into one of the biggest debates in all of sports: who's the greatest basketball player of all time? I'm so excited to have Sam here to talk about it with me. Thanks so much for having me, Alex. And you're absolutely right. This question always brings out some seriously strong feelings. Okay, so let's get right into it. For me, it has to be Michael Jordan. Six trips to the finals, six championships. That kind of perfection is just incredible. Oh man, the first thing that always pops into my head is that shot against the Cleveland Cavaliers back in '89. Jordan just rises, hangs in the air forever, and just sinks it.",
},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/example_output/VibeVoice-1.5B_output.wav",
},
],
}
],
]
# Process with apply_chat_template and output_labels=True for training
inputs = processor.apply_chat_template(
conversation,
tokenize=True,
return_dict=True,
processor_kwargs={"output_labels": True},
).to(model.device, model.dtype)
# Forward pass
outputs = model(**inputs, ddpm_batch_multiplier=2, num_diffusion_steps=2)
print(f"Total loss: {outputs.loss.item():.4f}")
# Backward pass
outputs.loss.backward()
Torch compile
The model can be compiled with torch.compile for faster inference. A few warmup runs are needed before the compiled model reaches full speed.
On an A100 with batch size 4, we observed a speed-up between compiled vs. non-compiled inference, see this script.
import os
import time
import torch
from transformers import AutoModelForTextToWaveform, AutoProcessor, CompileConfig
model_id = "microsoft/VibeVoice-7B-hf"
num_warmup = 5
max_new_tokens = 128
torch.set_float32_matmul_precision("high")
# Load processor + model
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForTextToWaveform.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto").eval()
# Prepare inputs
conversation = [
[
{
"role": "0",
"content": [
{"type": "text", "text": "VibeVoice is a novel framework for generating expressive audio."},
{
"type": "audio",
"url": "https://huggingface.co/datasets/bezzam/vibevoice_samples/resolve/main/realtime_model/vibevoice_tts_german.wav",
},
],
}
],
] * 4 # batch size 4
inputs = processor.apply_chat_template(
conversation, tokenize=True, return_dict=True, add_generation_prompt=True,
).to(model.device, model.dtype)
compile_config = CompileConfig(mode="default", dynamic=False)
generate_kwargs = dict(
**inputs,
max_new_tokens=max_new_tokens,
cache_implementation="static",
compile_config=compile_config,
)
# Warmup
print("Warming up...")
warmup_start = time.time()
with torch.inference_mode():
for _ in range(num_warmup):
torch.compiler.cudagraph_mark_step_begin()
_ = model.generate(**generate_kwargs)
torch.cuda.synchronize()
print(f"Warmup complete in {time.time() - warmup_start:.2f}s. Ready!")
# Apply model
with torch.inference_mode():
torch.compiler.cudagraph_mark_step_begin()
audio = model.generate(**generate_kwargs)
output_folder = f"{os.path.basename(model_id)}_compiled_output"
processor.save_audio(audio, output_folder)
print(f"Saved output to {output_folder}")
Training Details
Transformer-based Large Language Model (LLM) integrated with specialized acoustic and semantic tokenizers and a diffusion-based decoding head.
- LLM: Qwen2.5-7B for this release.
- Tokenizers:
- Acoustic Tokenizer: Based on a σ-VAE variant (proposed in LatentLM), with a mirror-symmetric encoder-decoder structure featuring 7 stages of modified Transformer blocks. Achieves 3200x downsampling from 24kHz input. Encoder/decoder components are ~340M parameters each.
- Semantic Tokenizer: Encoder mirrors the Acoustic Tokenizer's architecture (without VAE components). Trained with an ASR proxy task.
- Diffusion Head: Lightweight module (4 layers, ~123M parameters) conditioned on LLM hidden states. Predicts acoustic VAE features using a Denoising Diffusion Probabilistic Models (DDPM) process. Uses Classifier-Free Guidance (CFG) and DPM-Solver (and variants) during inference.
- Context Length: Trained with a curriculum increasing up to 32,768 tokens.
- Training Stages:
- Tokenizer Pre-training: Acoustic and Semantic tokenizers are pre-trained separately.
- VibeVoice Training: Pre-trained tokenizers are frozen; only the LLM and diffusion head parameters are trained. A curriculum learning strategy is used for input sequence length (4k -> 16K -> 32K). Text tokenizer not explicitly specified, but the LLM (Qwen2.5) typically uses its own. Audio is "tokenized" via the acoustic and semantic tokenizers.
Responsible Usage
Direct intended uses
The VibeVoice model is limited to research purpose use exploring highly realistic audio dialogue generation detailed in the tech report.
Out-of-scope uses
Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in any other way that is prohibited by MIT License. Use to generate any text transcript. Furthermore, this release is not intended or licensed for any of the following scenarios:
- Voice impersonation without explicit, recorded consent – cloning a real individual's voice for satire, advertising, ransom, social‑engineering, or authentication bypass.
- Disinformation or impersonation – creating audio presented as genuine recordings of real people or events.
- Real‑time or low‑latency voice conversion – telephone or video‑conference "live deep‑fake" applications.
- Unsupported language – the model is trained only on English and Chinese data; outputs in other languages are unsupported and may be unintelligible or offensive.
- Generation of background ambience, Foley, or music – VibeVoice is speech‑only and will not produce coherent non‑speech audio.
Risks and limitations
While efforts have been made to optimize it through various techniques, it may still produce outputs that are unexpected, biased, or inaccurate. VibeVoice inherits any biases, errors, or omissions produced by its base model (specifically, Qwen2.5 7b in this release). Potential for Deepfakes and Disinformation: High-quality synthetic speech can be misused to create convincing fake audio content for impersonation, fraud, or spreading disinformation. Users must ensure transcripts are reliable, check content accuracy, and avoid using generated content in misleading ways. Users are expected to use the generated content and to deploy the models in a lawful manner, in full compliance with all applicable laws and regulations in the relevant jurisdictions. It is best practice to disclose the use of AI when sharing AI-generated content. English and Chinese only: Transcripts in language other than English or Chinese may result in unexpected audio outputs. Non-Speech Audio: The model focuses solely on speech synthesis and does not handle background noise, music, or other sound effects. Overlapping Speech: The current model does not explicitly model or generate overlapping speech segments in conversations.
Recommendations
We do not recommend using VibeVoice in commercial or real-world applications without further testing and development. This model is intended for research and development purposes only. Please use responsibly.
To mitigate the risks of misuse, we have: Embedded an audible disclaimer (e.g. "This segment was generated by AI") automatically into every synthesized audio file. Added an imperceptible watermark to generated audio so third parties can verify VibeVoice provenance. Please see contact information at the end of this model card. Logged inference requests (hashed) for abuse pattern detection and publishing aggregated statistics quarterly. Users are responsible for sourcing their datasets legally and ethically. This may include securing appropriate rights and/or anonymizing data prior to use with VibeVoice. Users are reminded to be mindful of data privacy concerns.
Contact
This project was conducted by members of Microsoft Research. We welcome feedback and collaboration from our audience. If you have suggestions, questions, or observe unexpected/offensive behavior in our technology, please contact us at VibeVoice@microsoft.com. If the team receives reports of undesired behavior or identifies issues independently, we will update this repository with appropriate mitigations.
- Downloads last month
- 595