Conformer-Based Target-Speaker Automatic Speech Recognition For Single-Channel Audio

Abstract

We propose CONF-TSASR, a non-autoregressive end-toend time-frequency domain architecture for single-channel target-speaker automatic speech recognition (TS-ASR). The model consists of a TitaNet based speaker embedding module, a Conformer based masking as well as ASR modules. These modules are jointly optimized to transcribe a targetspeaker, while ignoring speech from other speakers. For training we use Connectionist Temporal Classification (CTC) loss and introduce a scale-invariant spectrogram reconstruction loss to encourage the model better separate the targetspeaker's spectrogram from mixture. We obtain state-of-theart target-speaker word error rate (TS-WER) on WSJ0-2mixextr (4.2%). Further, we report for the first time TS-WER on WSJ0-3mix-extr (12.4%), LibriSpeech2Mix (4.2%) and LibriSpeech3Mix (7.6%) datasets, establishing new benchmarks for TS-ASR. The proposed model will be open-sourced through NVIDIA NeMo toolkit.