Text-only domain adaptation for end-to-end ASR using integrated text-to-mel-spectrogram generator

TL;DR

The paper addresses domain adaptation for end-to-end ASR by enabling training on text-only data through an integrated text-to-spectrogram front-end that generates $80$-band mel spectrograms on the fly. A lightweight StyleGAN2-based Enhancer refines the synthetic spectrograms to reduce mismatch with real data, improving ASR performance without requiring external audio storage or vocoders. Across LibriSpeech, SLURP, and WSJ, the approach yields substantial gains in text-only adaptation, including up to a $41.2egin{smallmatrix}- ext{relative}rac{}{} ext{relative}\

Abstract

We propose an end-to-end Automatic Speech Recognition (ASR) system that can be trained on transcribed speech data, text-only data, or a mixture of both. The proposed model uses an integrated auxiliary block for text-based training. This block combines a non-autoregressive multi-speaker text-to-mel-spectrogram generator with a GAN-based enhancer to improve the spectrogram quality. The proposed system can generate a mel-spectrogram dynamically during training. It can be used to adapt the ASR model to a new domain by using text-only data from this domain. We demonstrate that the proposed training method significantly improves ASR accuracy compared to the system trained on transcribed speech only. It also surpasses cascade TTS systems with the vocoder in the adaptation quality and training speed.

Figures (2)

• Figure 1: ASR system with text-to-mel-spectrogram frontend
• Figure 2: Spectrogram Enhancer adversarial training setup