In this paper, we investigate the usage of large language models (LLMs) to improve the performance of competitive speech recognition systems. Different from traditional language models that focus on one single data domain, the rise of LLMs brings us the opportunity to push the limit of state-of-the-art ASR performance, and at the same time to achieve higher robustness and generalize effectively across multiple domains. Motivated by this, we propose a novel multi-stage approach to combine traditional language model re-scoring and LLM prompting. Specifically, the proposed method has two stages: the first stage uses a language model to re-score an N-best list of ASR hypotheses and run a confidence check; The second stage uses prompts to a LLM to perform ASR error correction on less confident results from the first stage. Our experimental results demonstrate the effectiveness of the proposed method by showing a 10% ~ 20% relative improvement in WER over a competitive ASR system -- across multiple test domains.
TorchAudio is an open-source audio and speech processing library built for PyTorch. It aims to accelerate the research and development of audio and speech technologies by providing well-designed, easy-to-use, and performant PyTorch components. Its contributors routinely engage with users to understand their needs and fulfill them by developing impactful features. Here, we survey TorchAudio's development principles and contents and highlight key features we include in its latest version (2.1): self-supervised learning pre-trained pipelines and training recipes, high-performance CTC decoders, speech recognition models and training recipes, advanced media I/O capabilities, and tools for performing forced alignment, multi-channel speech enhancement, and reference-less speech assessment. For a selection of these features, through empirical studies, we demonstrate their efficacy and show that they achieve competitive or state-of-the-art performance.
We propose a decoder-only language model, \textit{VoxtLM}, that can perform four tasks: speech recognition, speech synthesis, text generation, and speech continuation. VoxtLM integrates text vocabulary with discrete speech tokens from self-supervised speech features and uses special tokens to enable multitask learning. Compared to a single-task model, VoxtLM exhibits a significant improvement in speech synthesis, with improvements in both speech intelligibility from 28.9 to 5.6 and objective quality from 2.68 to 3.90. VoxtLM also improves speech generation and speech recognition performance over the single-task counterpart. VoxtLM is trained with publicly available data and training recipes and model checkpoints will be open-sourced to make fully reproducible work.
Fine-tuning large language models (LLMs) with domain-specific instructions has emerged as an effective method to enhance their domain-specific understanding. Yet, there is limited work that examines the core characteristics acquired during this process. In this study, we benchmark the fundamental characteristics learned by contact-center (CC) specific instruction fine-tuned LLMs with out-of-the-box (OOB) LLMs via probing tasks encompassing conversational, channel, and automatic speech recognition (ASR) properties. We explore different LLM architectures (Flan-T5 and Llama), sizes (3B, 7B, 11B, 13B), and fine-tuning paradigms (full fine-tuning vs PEFT). Our findings reveal remarkable effectiveness of CC-LLMs on the in-domain downstream tasks, with improvement in response acceptability by over 48% compared to OOB-LLMs. Additionally, we compare the performance of OOB-LLMs and CC-LLMs on the widely used SentEval dataset, and assess their capabilities in terms of surface, syntactic, and semantic information through probing tasks. Intriguingly, we note a relatively consistent performance of probing classifiers on the set of probing tasks. Our observations indicate that CC-LLMs, while outperforming their out-of-the-box counterparts, exhibit a tendency to rely less on encoding surface, syntactic, and semantic properties, highlighting the intricate interplay between domain-specific adaptation and probing task performance opening up opportunities to explore behavior of fine-tuned language models in specialized contexts.
Automatic Speech Recognition (ASR) in conversational settings presents unique challenges, including extracting relevant contextual information from previous conversational turns. Due to irrelevant content, error propagation, and redundancy, existing methods struggle to extract longer and more effective contexts. To address this issue, we introduce a novel Conversational ASR system, extending the Conformer encoder-decoder model with cross-modal conversational representation. Our approach leverages a cross-modal extractor that combines pre-trained speech and text models through a specialized encoder and a modal-level mask input. This enables the extraction of richer historical speech context without explicit error propagation. We also incorporate conditional latent variational modules to learn conversational level attributes such as role preference and topic coherence. By introducing both cross-modal and conversational representations into the decoder, our model retains context over longer sentences without information loss, achieving relative accuracy improvements of 8.8% and 23% on Mandarin conversation datasets HKUST and MagicData-RAMC, respectively, compared to the standard Conformer model.
Transformer-based models excel in speech recognition. Existing efforts to optimize Transformer inference, typically for long-context applications, center on simplifying attention score calculations. However, streaming speech recognition models usually process a limited number of tokens each time, making attention score calculation less of a bottleneck. Instead, the bottleneck lies in the linear projection layers of multi-head attention and feedforward networks, constituting a substantial portion of the model size and contributing significantly to computation, memory, and power usage. To address this bottleneck, we propose folding attention, a technique targeting these linear layers, significantly reducing model size and improving memory and power efficiency. Experiments on on-device Transformer-based streaming speech recognition models show that folding attention reduces model size (and corresponding memory consumption) by up to 24% and power consumption by up to 23%, all without compromising model accuracy or computation overhead.
Languages have long been described according to their perceived rhythmic attributes. The associated typologies are of interest in psycholinguistics as they partly predict newborns' abilities to discriminate between languages and provide insights into how adult listeners process non-native languages. Despite the relative success of rhythm metrics in supporting the existence of linguistic rhythmic classes, quantitative studies have yet to capture the full complexity of temporal regularities associated with speech rhythm. We argue that deep learning offers a powerful pattern-recognition approach to advance the characterization of the acoustic bases of speech rhythm. To explore this hypothesis, we trained a medium-sized recurrent neural network on a language identification task over a large database of speech recordings in 21 languages. The network had access to the amplitude envelopes and a variable identifying the voiced segments, assuming that this signal would poorly convey phonetic information but preserve prosodic features. The network was able to identify the language of 10-second recordings in 40% of the cases, and the language was in the top-3 guesses in two-thirds of the cases. Visualization methods show that representations built from the network activations are consistent with speech rhythm typologies, although the resulting maps are more complex than two separated clusters between stress and syllable-timed languages. We further analyzed the model by identifying correlations between network activations and known speech rhythm metrics. The findings illustrate the potential of deep learning tools to advance our understanding of speech rhythm through the identification and exploration of linguistically relevant acoustic feature spaces.
Most of the current speech data augmentation methods operate on either the raw waveform or the amplitude spectrum of speech. In this paper, we propose a novel speech data augmentation method called PhasePerturbation that operates dynamically on the phase spectrum of speech. Instead of statically rotating a phase by a constant degree, PhasePerturbation utilizes three dynamic phase spectrum operations, i.e., a randomization operation, a frequency masking operation, and a temporal masking operation, to enhance the diversity of speech data. We conduct experiments on wav2vec2.0 pre-trained ASR models by fine-tuning them with the PhasePerturbation augmented TIMIT corpus. The experimental results demonstrate 10.9\% relative reduction in the word error rate (WER) compared with the baseline model fine-tuned without any augmentation operation. Furthermore, the proposed method achieves additional improvements (12.9\% and 15.9\%) in WER by complementing the Vocal Tract Length Perturbation (VTLP) and the SpecAug, which are both amplitude spectrum-based augmentation methods. The results highlight the capability of PhasePerturbation to improve the current amplitude spectrum-based augmentation methods.
Second pass rescoring is a critical component of competitive automatic speech recognition (ASR) systems. Large language models have demonstrated their ability in using pre-trained information for better rescoring of ASR hypothesis. Discriminative training, directly optimizing the minimum word-error-rate (MWER) criterion typically improves rescoring. In this study, we propose and explore several discriminative fine-tuning schemes for pre-trained LMs. We propose two architectures based on different pooling strategies of output embeddings and compare with probability based MWER. We conduct detailed comparisons between pre-trained causal and bidirectional LMs in discriminative settings. Experiments on LibriSpeech demonstrate that all MWER training schemes are beneficial, giving additional gains upto 8.5\% WER. Proposed pooling variants achieve lower latency while retaining most improvements. Finally, our study concludes that bidirectionality is better utilized with discriminative training.
ASR models often suffer from a long-form deletion problem where the model predicts sequential blanks instead of words when transcribing a lengthy audio (in the order of minutes or hours). From the perspective of a user or downstream system consuming the ASR results, this behavior can be perceived as the model "being stuck", and potentially make the product hard to use. One of the culprits for long-form deletion is training-test data mismatch, which can happen even when the model is trained on diverse and large-scale data collected from multiple application domains. In this work, we introduce a novel technique to simultaneously model different groups of speakers in the audio along with the standard transcript tokens. Speakers are grouped as primary and non-primary, which connects the application domains and significantly alleviates the long-form deletion problem. This improved model neither needs any additional training data nor incurs additional training or inference cost.