Papers
arxiv:2505.03739

VITA-Audio: Fast Interleaved Cross-Modal Token Generation for Efficient Large Speech-Language Model

Published on May 6
Β· Submitted by shenyunhang on May 7
Authors:
,
,
,
,
,
,
,
,
,
,
,
,
,

Abstract

With the growing requirement for natural human-computer interaction, speech-based systems receive increasing attention as speech is one of the most common forms of daily communication. However, the existing speech models still experience high latency when generating the first audio token during streaming, which poses a significant bottleneck for deployment. To address this issue, we propose VITA-Audio, an end-to-end large speech model with fast audio-text token generation. Specifically, we introduce a lightweight Multiple Cross-modal Token Prediction (MCTP) module that efficiently generates multiple audio tokens within a single model forward pass, which not only accelerates the inference but also significantly reduces the latency for generating the first audio in streaming scenarios. In addition, a four-stage progressive training strategy is explored to achieve model acceleration with minimal loss of speech quality. To our knowledge, VITA-Audio is the first multi-modal large language model capable of generating audio output during the first forward pass, enabling real-time conversational capabilities with minimal latency. VITA-Audio is fully reproducible and is trained on open-source data only. Experimental results demonstrate that our model achieves an inference speedup of 3~5x at the 7B parameter scale, but also significantly outperforms open-source models of similar model size on multiple benchmarks for automatic speech recognition (ASR), text-to-speech (TTS), and spoken question answering (SQA) tasks.

Community

Paper submitter
β€’
edited 2 days ago

image.png

✨ Highlights

  • Low Latency. VITA-Audio is the first end-to-end speech model capable of generating audio during the initial forward pass. By utilizing a set of 32 prefill tokens, VITA-Audio reduces the time required to generate the first audio token chunk from 236 ms to just 53 ms.
  • Fast Inference. VITA-Audio achieves an inference speedup of 3-5x at the 7B parameter scale.
  • Open Source. VITA-Audio is trained on open-source data only, consisting of 200k hours of publicly available audio.
  • Strong Performance. VITA-Audio achieves competitive results on ASR, TTS, and SQA benchmarks among cutting-edge models under 7B parameters.

πŸ“Œ Exhibition

Inference Acceleration

Model inference speed under different inference modes.

demogif second_gif

Time to Generate the First Audio Segment In Streaming Inference

first audio generate time

Generated Audio Case


To be or not to be--to live intensely and richly,
merely to exist, that depends on ourselves. Let widen and intensify our relations.
While we live, let live!


The hair has been so little, don't think about it, go to bed early, for your hair. Good night!

πŸ“ˆ Experimental Results

  • Comparison of Spoken Question Answering.

image

  • Comparison of Text to Speech.

image

  • Comparison of Automatic Speech Recognition.

image

image

  • Effectiveness of Inference Acceleration.

image

Image

Sign up or log in to comment

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2505.03739 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2505.03739 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2505.03739 in a Space README.md to link it from this page.

Collections including this paper 3