The possibility of dynamically modifying the computational load of neural models at inference time is crucial for on-device processing, where computational power is limited and time-varying. Established approaches for neural model compression exist, but they provide architecturally static models. In this paper, we investigate the use of early-exit architectures, that rely on intermediate exit branches, applied to large-vocabulary speech recognition. This allows for the development of dynamic models that adjust their computational cost to the available resources and recognition performance. Unlike previous works, besides using pre-trained backbones we also train the model from scratch with an early-exit architecture. Experiments on public datasets show that early-exit architectures from scratch not only preserve performance levels when using fewer encoder layers, but also improve task accuracy as compared to using single-exit models or using pre-trained models. Additionally, we investigate an exit selection strategy based on posterior probabilities as an alternative to frame-based entropy.
Soft random sampling (SRS) is a simple yet effective approach for efficient training of large-scale deep neural networks when dealing with massive data. SRS selects a subset uniformly at random with replacement from the full data set in each epoch. In this paper, we conduct a theoretical and empirical analysis of SRS. First, we analyze its sampling dynamics including data coverage and occupancy. Next, we investigate its convergence with non-convex objective functions and give the convergence rate. Finally, we provide its generalization performance. We empirically evaluate SRS for image recognition on CIFAR10 and automatic speech recognition on Librispeech and an in-house payload dataset to demonstrate its effectiveness. Compared to existing coreset-based data selection methods, SRS offers a better accuracy-efficiency trade-off. Especially on real-world industrial scale data sets, it is shown to be a powerful training strategy with significant speedup and competitive performance with almost no additional computing cost.
Advancements in deep neural networks have allowed automatic speech recognition (ASR) systems to attain human parity on several publicly available clean speech datasets. However, even state-of-the-art ASR systems experience performance degradation when confronted with adverse conditions, as a well-trained acoustic model is sensitive to variations in the speech domain, e.g., background noise. Intuitively, humans address this issue by relying on their linguistic knowledge: the meaning of ambiguous spoken terms is usually inferred from contextual cues thereby reducing the dependency on the auditory system. Inspired by this observation, we introduce the first open-source benchmark to utilize external large language models (LLMs) for ASR error correction, where N-best decoding hypotheses provide informative elements for true transcription prediction. This approach is a paradigm shift from the traditional language model rescoring strategy that can only select one candidate hypothesis as the output transcription. The proposed benchmark contains a novel dataset, HyPoradise (HP), encompassing more than 334,000 pairs of N-best hypotheses and corresponding accurate transcriptions across prevalent speech domains. Given this dataset, we examine three types of error correction techniques based on LLMs with varying amounts of labeled hypotheses-transcription pairs, which gains a significant word error rate (WER) reduction. Experimental evidence demonstrates the proposed technique achieves a breakthrough by surpassing the upper bound of traditional re-ranking based methods. More surprisingly, LLM with reasonable prompt and its generative capability can even correct those tokens that are missing in N-best list. We make our results publicly accessible for reproducible pipelines with released pre-trained models, thus providing a new evaluation paradigm for ASR error correction with LLMs.
Deep biasing for the Transducer can improve the recognition performance of rare words or contextual entities, which is essential in practical applications, especially for streaming Automatic Speech Recognition (ASR). However, deep biasing with large-scale rare words remains challenging, as the performance drops significantly when more distractors exist and there are words with similar grapheme sequences in the bias list. In this paper, we combine the phoneme and textual information of rare words in Transducers to distinguish words with similar pronunciation or spelling. Moreover, the introduction of training with text-only data containing more rare words benefits large-scale deep biasing. The experiments on the LibriSpeech corpus demonstrate that the proposed method achieves state-of-the-art performance on rare word error rate for different scales and levels of bias lists.
This paper presents an end-to-end model designed to improve automatic speech recognition (ASR) for a particular speaker in a crowded, noisy environment. The model utilizes a single-channel speech enhancement module that isolates the speaker's voice from background noise, along with an ASR module. Through this approach, the model is able to decrease the word error rate (WER) of ASR from 80% to 26.4%. Typically, these two components are adjusted independently due to variations in data requirements. However, speech enhancement can create anomalies that decrease ASR efficiency. By implementing a joint fine-tuning strategy, the model can reduce the WER from 26.4% in separate tuning to 14.5% in joint tuning.
Recently Whisper has approached human-level robustness and accuracy in English automatic speech recognition (ASR), while in minor language and mixed language speech recognition, there remains a compelling need for further improvement. In this work, we present the impressive results of Whisper-MCE, our finetuned Whisper model, which was trained using our self-collected dataset, Mixed Cantonese and English audio dataset (MCE). Meanwhile, considering word error rate (WER) poses challenges when it comes to evaluating its effectiveness in minor language and mixed-language contexts, we present a novel rating mechanism. By comparing our model to the baseline whisper-large-v2 model, we demonstrate its superior ability to accurately capture the content of the original audio, achieve higher recognition accuracy, and exhibit faster recognition speed. Notably, our model outperforms other existing models in the specific task of recognizing mixed language.
Deep neural networks (DNNs) exhibit superior performance in various machine learning tasks, e.g., image classification, speech recognition, biometric recognition, object detection, etc. However, it is essential to analyze their sensitivity to parameter perturbations before deploying them in real-world applications. In this work, we assess the sensitivity of DNNs against perturbations to their weight and bias parameters. The sensitivity analysis involves three DNN architectures (VGG, ResNet, and DenseNet), three types of parameter perturbations (Gaussian noise, weight zeroing, and weight scaling), and two settings (entire network and layer-wise). We perform experiments in the context of iris presentation attack detection and evaluate on two publicly available datasets: LivDet-Iris-2017 and LivDet-Iris-2020. Based on the sensitivity analysis, we propose improved models simply by perturbing parameters of the network without undergoing training. We further combine these perturbed models at the score-level and at the parameter-level to improve the performance over the original model. The ensemble at the parameter-level shows an average improvement of 43.58% on the LivDet-Iris-2017 dataset and 9.25% on the LivDet-Iris-2020 dataset. The source code is available at https://github.com/redwankarimsony/WeightPerturbation-MSU.
The performance of acoustic models degrades notably in noisy environments. Speech enhancement (SE) can be used as a front-end strategy to aid automatic speech recognition (ASR) systems. However, existing training objectives of SE methods are not fully effective at integrating speech-text and noisy-clean paired data for training toward unseen ASR systems. In this study, we propose a general denoising framework, D4AM, for various downstream acoustic models. Our framework fine-tunes the SE model with the backward gradient according to a specific acoustic model and the corresponding classification objective. In addition, our method aims to consider the regression objective as an auxiliary loss to make the SE model generalize to other unseen acoustic models. To jointly train an SE unit with regression and classification objectives, D4AM uses an adjustment scheme to directly estimate suitable weighting coefficients rather than undergoing a grid search process with additional training costs. The adjustment scheme consists of two parts: gradient calibration and regression objective weighting. The experimental results show that D4AM can consistently and effectively provide improvements to various unseen acoustic models and outperforms other combination setups. Specifically, when evaluated on the Google ASR API with real noisy data completely unseen during SE training, D4AM achieves a relative WER reduction of 24.65% compared with the direct feeding of noisy input. To our knowledge, this is the first work that deploys an effective combination scheme of regression (denoising) and classification (ASR) objectives to derive a general pre-processor applicable to various unseen ASR systems. Our code is available at https://github.com/ChangLee0903/D4AM.
Discrete audio representation, aka audio tokenization, has seen renewed interest driven by its potential to facilitate the application of text language modeling approaches in audio domain. To this end, various compression and representation-learning based tokenization schemes have been proposed. However, there is limited investigation into the performance of compression-based audio tokens compared to well-established mel-spectrogram features across various speaker and speech related tasks. In this paper, we evaluate compression based audio tokens on three tasks: Speaker Verification, Diarization and (Multi-lingual) Speech Recognition. Our findings indicate that (i) the models trained on audio tokens perform competitively, on average within $1\%$ of mel-spectrogram features for all the tasks considered, and do not surpass them yet. (ii) these models exhibit robustness for out-of-domain narrowband data, particularly in speaker tasks. (iii) audio tokens allow for compression to 20x compared to mel-spectrogram features with minimal loss of performance in speech and speaker related tasks, which is crucial for low bit-rate applications, and (iv) the examined Residual Vector Quantization (RVQ) based audio tokenizer exhibits a low-pass frequency response characteristic, offering a plausible explanation for the observed results, and providing insight for future tokenizer designs.
This study focuses on the development of Indonesian Automatic Speech Recognition (ASR) using the XLSR-53 pre-trained model, the XLSR stands for cross-lingual speech representations. The use of this XLSR-53 pre-trained model is to significantly reduce the amount of training data in non-English languages required to achieve a competitive Word Error Rate (WER). The total amount of data used in this study is 24 hours, 18 minutes, and 1 second: (1) TITML-IDN 14 hours and 31 minutes; (2) Magic Data 3 hours and 33 minutes; and (3) Common Voice 6 hours, 14 minutes, and 1 second. With a WER of 20%, the model built in this study can compete with similar models using the Common Voice dataset split test. WER can be decreased by around 8% using a language model, resulted in WER from 20% to 12%. Thus, the results of this study have succeeded in perfecting previous research in contributing to the creation of a better Indonesian ASR with a smaller amount of data.