While Speech Foundation Models (SFMs) excel in various speech tasks, their performance for low-resource tasks such as child Automatic Speech Recognition (ASR) is hampered by limited pretraining data. To address this, we explore different model merging techniques to leverage knowledge from models trained on larger, more diverse speech corpora. This paper also introduces Selective Attention (SA) Merge, a novel method that selectively merges task vectors from attention matrices to enhance SFM performance on low-resource tasks. Experiments on the MyST database show significant reductions in relative word error rate of up to 14%, outperforming existing model merging and data augmentation techniques. By combining data augmentation techniques with SA Merge, we achieve a new state-of-the-art WER of 8.69 on the MyST database for the Whisper-small model, highlighting the potential of SA Merge for improving low-resource ASR.

This paper presents a novel dataset (CORAAL QA) and framework for audio question-answering from long audio recordings containing spontaneous speech. The dataset introduced here provides sets of questions that can be factually answered from short spans of a long audio files (typically 30min to 1hr) from the Corpus of Regional African American Language. Using this dataset, we divide the audio recordings into 60 second segments, automatically transcribe each segment, and use PLDA scoring of BERT-based semantic embeddings to rank the relevance of ASR transcript segments in answering the target question. In order to improve this framework through data augmentation, we use large language models including ChatGPT and Llama 2 to automatically generate further training examples and show how prompt engineering can be optimized for this process. By creatively leveraging knowledge from large-language models, we achieve state-of-the-art question-answering performance in this information retrieval task.

Non-autoregressive automatic speech recognition (NASR) models have gained attention due to their parallelism and fast inference. The encoder-based NASR, e.g. connectionist temporal classification (CTC), can be initialized from the speech foundation models (SFM) but does not account for any dependencies among intermediate tokens. The encoder-decoder-based NASR, like CTC alignment-based single-step non-autoregressive transformer (CASS-NAT), can mitigate the dependency problem but is not able to efficiently integrate SFM. Inspired by the success of recent work of speech-text joint pre-training with a shared transformer encoder, we propose a new encoder-based NASR, UniEnc-CASSNAT, to combine the advantages of CTC and CASS-NAT. UniEnc-CASSNAT consists of only an encoder as the major module, which can be the SFM. The encoder plays the role of both the CASS-NAT encoder and decoder by two forward passes. The first pass of the encoder accepts the speech signal as input, while the concatenation of the speech signal and the token-level acoustic embedding is used as the input for the second pass. Examined on the Librispeech 100h, MyST, and Aishell1 datasets, the proposed UniEnc-CASSNAT achieves state-of-the-art NASR results and is better or comparable to CASS-NAT with only an encoder and hence, fewer model parameters. Our codes1are publicly available.