Muhammad Shakeel

Yui Sudo, Muhammad Shakeel, Brian Yan et al.

The network architecture of end-to-end (E2E) automatic speech recognition (ASR) can be classified into several models, including connectionist temporal classification (CTC), recurrent neural network transducer (RNN-T), attention mechanism, and non-autoregressive mask-predict models. Since each of these network architectures has pros and cons, a typical use case is to switch these separate models depending on the application requirement, resulting in the increased overhead of maintaining all models. Several methods for integrating two of these complementary models to mitigate the overhead issue have been proposed; however, if we integrate more models, we will further benefit from these complementary models and realize broader applications with a single system. This paper proposes four-decoder joint modeling (4D) of CTC, attention, RNN-T, and mask-predict, which has the following three advantages: 1) The four decoders are jointly trained so that they can be easily switched depending on the application scenarios. 2) Joint training may bring model regularization and improve the model robustness thanks to their complementary properties. 3) Novel one-pass joint decoding methods using CTC, attention, and RNN-T further improves the performance. The experimental results showed that the proposed model consistently reduced the WER.

Yifan Peng, Yui Sudo, Muhammad Shakeel et al. · nvidia

There has been an increasing interest in large speech models that can perform multiple tasks in a single model. Such models usually adopt an encoder-decoder or decoder-only architecture due to their popularity and good performance in many domains. However, autoregressive models can be slower during inference compared to non-autoregressive models and also have potential risks of hallucination. Though prior studies observed promising results of non-autoregressive models for certain tasks at small scales, it remains unclear if they can be scaled to speech-to-text generation in diverse languages and tasks. Inspired by the Open Whisper-style Speech Model (OWSM) project, we propose OWSM-CTC, a novel encoder-only speech foundation model based on Connectionist Temporal Classification (CTC). It is trained on 180k hours of public audio data for multilingual automatic speech recognition (ASR), speech translation (ST), and language identification (LID). Compared to encoder-decoder OWSM, our OWSM-CTC achieves competitive results on ASR and up to 24% relative improvement on ST, while it is more robust and 3 to 4 times faster for inference. OWSM-CTC also improves the long-form ASR result with 20x speed-up. We will publicly release our code, pre-trained model, and training logs to promote open science in speech foundation models.

Yui Sudo, Yosuke Fukumoto, Muhammad Shakeel et al. · nvidia

Deep biasing (DB) enhances the performance of end-to-end automatic speech recognition (E2E-ASR) models for rare words or contextual phrases using a bias list. However, most existing methods treat bias phrases as sequences of subwords in a predefined static vocabulary. This naive sequence decomposition produces unnatural token patterns, significantly lowering their occurrence probability. More advanced techniques address this problem by expanding the vocabulary with additional modules, including the external language model shallow fusion or rescoring. However, they result in increasing the workload due to the additional modules. This paper proposes a dynamic vocabulary where bias tokens can be added during inference. Each entry in a bias list is represented as a single token, unlike a sequence of existing subword tokens. This approach eliminates the need to learn subword dependencies within the bias phrases. This method is easily applied to various architectures because it only expands the embedding and output layers in common E2E-ASR architectures. Experimental results demonstrate that the proposed method improves the bias phrase WER on English and Japanese datasets by 3.1 -- 4.9 points compared with the conventional DB method.

Muhammad Shakeel, Yui Sudo, Yifan Peng et al. · nvidia

This paper presents a unified multi-speaker encoder (UME), a novel architecture that jointly learns representations for speaker diarization (SD), speech separation (SS), and multi-speaker automatic speech recognition (ASR) tasks using a shared speech foundational encoder. We leverage the hidden representations from multiple layers of UME as a residual weighted-sum encoding (RWSE) to effectively use information from different semantic levels, contributing to bottom-up alignment between tasks. This joint training approach captures the inherent interdependencies among the tasks, enhancing overall performance on overlapping speech data. Our evaluations demonstrate that UME substantially improves over the single-task baselines dedicated to SD, SS, and multi-speaker ASR on LibriMix evaluation sets. Notably, for SD, UME outperforms the previous studies, achieving diarization error rates of 1.37% and 2.29% on Libri2Mix and Libri3Mix evaluation sets, respectively.

Yui Sudo, Yosuke Fukumoto, Muhammad Shakeel et al. · nvidia

Contextual biasing (CB) improves automatic speech recognition for rare and unseen phrases. Recent studies have introduced dynamic vocabulary, which represents context phrases as expandable tokens in autoregressive (AR) models. This method improves CB accuracy but with slow inference speed. While dynamic vocabulary can be applied to non-autoregressive (NAR) models, such as connectionist temporal classification (CTC), the conditional independence assumption fails to capture dependencies between static and dynamic tokens. This paper proposes DYNAC (Dynamic Vocabulary-based NAR Contextualization), a self-conditioned CTC method that integrates dynamic vocabulary into intermediate layers. Conditioning the encoder on dynamic vocabulary, DYNAC effectively captures dependencies between static and dynamic tokens while reducing the real-time factor (RTF). Experimental results show that DYNAC reduces RTF by 81% with a 0.1-point degradation in word error rate on the LibriSpeech 960 test-clean set.

Muhammad Shakeel, Yui Sudo, Yifan Peng et al.

Contextualized end-to-end automatic speech recognition has been an active research area, with recent efforts focusing on the implicit learning of contextual phrases based on the final loss objective. However, these approaches ignore the useful contextual knowledge encoded in the intermediate layers. We hypothesize that employing explicit biasing loss as an auxiliary task in the encoder intermediate layers may better align text tokens or audio frames with the desired objectives. Our proposed intermediate biasing loss brings more regularization and contextualization to the network. Our method outperforms a conventional contextual biasing baseline on the LibriSpeech corpus, achieving a relative improvement of 22.5% in biased word error rate (B-WER) and up to 44% compared to the non-contextual baseline with a biasing list size of 100. Moreover, employing RNN-transducer-driven joint decoding further reduces the unbiased word error rate (U-WER), resulting in a more robust network.

Muhammad Shakeel, Yui Sudo, Yifan Peng et al. · nvidia

End-to-end (E2E) automatic speech recognition (ASR) can operate in two modes: streaming and non-streaming, each with its pros and cons. Streaming ASR processes the speech frames in real-time as it is being received, while non-streaming ASR waits for the entire speech utterance; thus, professionals may have to operate in either mode to satisfy their application. In this work, we present joint optimization of streaming and non-streaming ASR based on multi-decoder and knowledge distillation. Primarily, we study 1) the encoder integration of these ASR modules, followed by 2) separate decoders to make the switching mode flexible, and enhancing performance by 3) incorporating similarity-preserving knowledge distillation between the two modular encoders and decoders. Evaluation results show 2.6%-5.3% relative character error rate reductions (CERR) on CSJ for streaming ASR, and 8.3%-9.7% relative CERRs for non-streaming ASR within a single model compared to multiple standalone modules.

Yui Sudo, Muhammad Shakeel, Yosuke Fukumoto et al.

End-to-end (E2E) automatic speech recognition (ASR) methods exhibit remarkable performance. However, since the performance of such methods is intrinsically linked to the context present in the training data, E2E-ASR methods do not perform as desired for unseen user contexts (e.g., technical terms, personal names, and playlists). Thus, E2E-ASR methods must be easily contextualized by the user or developer. This paper proposes an attention-based contextual biasing method that can be customized using an editable phrase list (referred to as a bias list). The proposed method can be trained effectively by combining a bias phrase index loss and special tokens to detect the bias phrases in the input speech data. In addition, to improve the contextualization performance during inference further, we propose a bias phrase boosted (BPB) beam search algorithm based on the bias phrase index probability. Experimental results demonstrate that the proposed method consistently improves the word error rate and the character error rate of the target phrases in the bias list on both the Librispeech-960 (English) and our in-house (Japanese) dataset, respectively.

Muhammad Shakeel, Katsutoshi Itoyama, Kenji Nishida et al.

We describe a novel metric-based learning approach that introduces a multimodal framework and uses deep audio and geophone encoders in siamese configuration to design an adaptable and lightweight supervised model. This framework eliminates the need for expensive data labeling procedures and learns general-purpose representations from low multisensory data obtained from omnipresent sensing systems. These sensing systems provide numerous applications and various use cases in activity recognition tasks. Here, we intend to explore the human footstep movements from indoor environments and analyze representations from a small self-collected dataset of acoustic and vibration-based sensors. The core idea is to learn plausible similarities between two sensory traits and combining representations from audio and geophone signals. We present a generalized framework to learn embeddings from temporal and spatial features extracted from audio and geophone signals. We then extract the representations in a shared space to maximize the learning of a compatibility function between acoustic and geophone features. This, in turn, can be used effectively to carry out a classification task from the learned model, as demonstrated by assigning high similarity to the pairs with a human footstep movement and lower similarity to pairs containing no footstep movement. Performance analyses show that our proposed multimodal framework achieves a 19.99\% accuracy increase (in absolute terms) and avoided overfitting on the evaluation set when the training samples were increased from 200 pairs to just 500 pairs while satisfactorily learning the audio and geophone representations. Our results employ a metric-based contrastive learning approach for multi-sensor data to mitigate the impact of data scarcity and perform human movement identification with limited data size.