While large language models (LLMs) have been applied to automatic speech recognition (ASR), the task of making the model streamable remains a challenge. This paper proposes a novel model architecture, Transducer-Llama, that integrates LLMs into a Factorized Transducer (FT) model, naturally enabling streaming capabilities. Furthermore, given that the large vocabulary of LLMs can cause data sparsity issue and increased training costs for spoken language systems, this paper introduces an efficient vocabulary adaptation technique to align LLMs with speech system vocabularies. The results show that directly optimizing the FT model with a strong pre-trained LLM-based predictor using the RNN-T loss yields some but limited improvements over a smaller pre-trained LM predictor. Therefore, this paper proposes a weak-to-strong LM swap strategy, using a weak LM predictor during RNN-T loss training and then replacing it with a strong LLM. After LM replacement, the minimum word error rate (MWER) loss is employed to finetune the integration of the LLM predictor with the Transducer-Llama model. Experiments on the LibriSpeech and large-scale multi-lingual LibriSpeech corpora show that the proposed streaming Transducer-Llama approach gave a 17% relative WER reduction (WERR) over a strong FT baseline and a 32% WERR over an RNN-T baseline.

We introduce DAS (Domain Adaptation with Synthetic data), a novel domain adaptation framework for pre-trained ASR model, designed to efficiently adapt to various language-defined domains without requiring any real data. In particular, DAS first prompts large language models (LLMs) to generate domain-specific texts before converting these texts to speech via text-to-speech technology. The synthetic data is used to finetune Whisper with Low-Rank Adapters (LoRAs) for targeted domains such as music, weather, and sports. We introduce a novel one-pass decoding strategy that merges predictions from multiple LoRA adapters efficiently during the auto-regressive text generation process. Experimental results show significant improvements, reducing the Word Error Rate (WER) by 10% to 17% across all target domains compared to the original model, with minimal performance regression in out-of-domain settings −1% on Librispeech test sets. We also demonstrate that DAS operates efficiently during inference, introducing an additional 9% increase in Real Time Factor (RTF) compared to the original model when inferring with three LoRA adapters.

Large language models (LLMs) have proven themselves highly flexible, able to solve a wide range of generative tasks, such as abstractive summarization and open-ended question answering. In this paper we extend the capabilities of LLM by directly attaching a small audio encoder allowing it to perform speech recognition. By directly prepending a sequence of audio embeddings to the text token embeddings, the LLM can be converted to an automatic speech recognition (ASR) system, and be used in the exact same manner as its textual counterpart. Experiments on Multilingual LibriSpeech (MLS) show that incorporating a conformer encoder into the open sourced LLaMA-7B allows it to outperform monolingual baselines by 18% relatively in WER and perform multilingual speech recognition, despite LLaMA being trained overwhelmingly on English text. Furthermore, we perform ablation studies to investigate whether the LLM can be completely frozen during training to maintain its original capabilities, scaling up the audio encoder, and increasing the audio encoder striding to generate fewer embeddings. The results from these studies show that multilingual ASR is possible even when the LLM is frozen, or when strides of almost 1 second are used in the audio encoder opening up the possibility for LLMs to operate on long-form audio.

The internal language model (ILM) of the neural transducer has been widely studied. In most prior work, it is mainly used for estimating the ILM score and is subsequently subtracted during inference to facilitate improved integration with external language models. Recently, various of factorized transducer models have been proposed, which explicitly embrace a standalone internal language model for non-blank token prediction. However, even with the adoption of factorized transducer models, limited improvement has been observed compared to shallow fusion. In this paper, we propose a novel ILM training and decoding strategy for factorized transducer models, which effectively combines the blank, acoustic and ILM scores. Our experiments show a 17% relative improvement over the standard decoding method when utilizing a well-trained ILM and the proposed decoding strategy on LibriSpeech datasets. Furthermore, when compared to a strong RNN-T baseline enhanced with external LM fusion, the proposed model yields a 5.5% relative improvement on general-sets and an 8.9% WER reduction for rare words. The proposed model can achieve superior performance without relying on external language models, rendering it highly efficient for production use-cases. To further improve the performance, we propose a novel and memory-efficient ILM-fusion-aware minimum word error rate (MWER) training method which improves ILM integration significantly.

Neural network pruning offers an effective method for compressing a multilingual automatic speech recognition (ASR) model with minimal performance loss. However, it entails several rounds of pruning and re-training needed to be run for each language. In this work, we propose the use of an adaptive masking approach in two scenarios for pruning a multilingual ASR model efficiently, each resulting in sparse monolingual models or a sparse multilingual model (named as Dynamic ASR Pathways). Our approach dynamically adapts the subnetwork, avoiding premature decisions about a fixed sub-network structure. We show that our approach outperforms existing pruning methods when targeting sparse monolingual models. Further, we illustrate that Dynamic ASR Pathways jointly discovers and trains better sub-networks (pathways) of a single multilingual model by adapting from different sub-network initializations, thereby reducing the need for language-specific pruning.

This paper introduces a fast-slow encoder based transducer with streaming deliberation for end-to-end automatic speech recognition. We aim to improve the recognition accuracy of the fast-slow encoder based transducer while keeping its latency low by integrating a streaming deliberation model. Specifically, the deliberation model leverages partial hypotheses from the streaming fast encoder and implicitly learns to correct recognition errors. We modify the parallel beam search algorithm for fast-slow encoder based transducer to be efficient and compatible with the deliberation model. In addition, the deliberation model is designed to process streaming data. To further improve the deliberation performance, a simple text augmentation approach is explored. We also compare LSTM and Conformer models for encoding partial hypotheses. Experiments on Librispeech and in-house data show relative WER reductions (WERRs) from 3% to 5% with a slight increase in model size and negligible extra token emission latency compared with fast-slow encoder based transducer. The deliberation model also reduces rare WERs by 3-7% on large-scale in-house data. Compared with vanilla neural transducers, the proposed deliberation model together with fast-slow encoder based transducer obtains relative 10-11% WERRs on Librispeech and around relative 6% WERR on in-house data with smaller emission delays.

While end-to-end models have shown great success on the Automatic Speech Recognition task, performance degrades severely when target sentences are long-form. The previous proposed methods, (partial) overlapping inference are shown to be effective on long-form decoding. For both methods, word error rate (WER) decreases monotonically when over-lapping percentage decreases. Setting aside computational cost, the setup with 50% overlapping during inference can achieve the best performance. However, a lower overlapping percentage has an advantage of fast inference speed. In this paper, we first conduct comprehensive experiments comparing overlapping inference and partial overlapping inference with various configurations. We then propose Voice-Activity-Detection Overlapping Inference to provide a trade-off between WER and computation cost. Results show that the pro-posed method can achieve a 20% relative computation cost reduction on Librispeech and Microsoft Speech Language Translation long-form corpus while maintaining the WER performance when comparing to the best performing over-lapping inference algorithm. We also propose Soft-Match to compensate for similar words misaligned problem.

Accents mismatching is a critical problem for end-to-end ASR. This paper aims to address this problem by building an accent-robust RNN-T system with domain adversarial training (DAT). We unveil the magic behind DAT and provide, for the first time, a theoretical guarantee that DAT learns accent-invariant representations. We also prove that performing the gradient reversal in DAT is equivalent to minimizing the Jensen-Shannon divergence between domain output distributions. Motivated by the proof of equivalence, we introduce reDAT, a novel technique based on DAT, which relabels data using either unsupervised clustering or soft labels. Experiments on 23K hours of multi-accent data show that DAT achieves competitive results over accent-specific baselines on both native and non-native English accents but up to 13% relative WER reduction on unseen accents; our reDAT yields further improvements over DAT by 3% and 8% relatively on non-native accents of American and British English.

The effects of speaking-style variability on automatic speaker verification were investigated using the UCLA Speaker Variability database which comprises multiple speaking styles per speaker. An x-vector/PLDA (probabilistic linear discriminant analysis) system was trained with the SRE and Switchboard databases with standard augmentation techniques and evaluated with utterances from the UCLA database. The equal error rate (EER) was low when enrollment and test utterances were of the same style (e.g., 0.98% and 0.57% for read and conversational speech, respectively), but it increased substantially when styles were mismatched between enrollment and test utterances. For instance, when enrolled with conversation utterances, the EER increased to 3.03%, 2.96% and 22.12% when tested on read, narrative, and pet-directed speech, respectively. To reduce the effect of style mismatch, we propose an entropy-based variable frame rate technique to artificially generate style-normalized representations for PLDA adaptation. The proposed system significantly improved performance. In the aforementioned conditions, the EERs improved to 2.69% (conversation -- read), 2.27% (conversation -- narrative), and 18.75% (pet-directed -- read). Overall, the proposed technique performed comparably to multi-style PLDA adaptation without the need for training data in different speaking styles per speaker.

In this work, we propose a novel and efficient minimum word error rate (MWER) training method for RNN-Transducer (RNN-T).Unlike previous work on this topic, which performs on-the-fly limited-size beam-search decoding and generates alignment scores for expected edit-distance computation, in our proposed method, we re-calculate and sum scores of all the possible alignments for each hypothesis in N-best lists.The hypothesis probability scores and back-propagated gradients are calculated efficiently using the forward-backward algorithm.Moreover, the proposed method allows us to decouple the decoding and training processes, and thus we can perform offline parallel-decoding and MWER training for each subset iteratively.Experimental results show that this proposed semi-on-the-fly method can speed up the on-the-fly method by 6 times and result in a similar WER improvement (3.6%) over a baseline RNN-T model.The proposed MWER training can also effectively reduce high-deletion errors (9.2% WER-reduction) introduced by RNN-T models when EOS is added for endpointer.Further improvement can be achieved if we use a proposed RNN-T rescoring method to re-rank hypotheses and use external RNN-LM to perform additional rescoring.The best system achieves a 5% relative improvement on an English test-set of real far-field recordings and a 11.6% WER reduction on music-domain utterances.

Attention-based sequence-to-sequence models for speech recognition jointly train an acoustic model, language model (LM), and alignment mechanism using a single neural network and require only parallel audio-text pairs. Thus, the language model component of the end-to-end model is only trained on transcribed audio-text pairs, which leads to performance degradation especially on rare words. While there have been a variety of work that look at incorporating an external LM trained on text-only data into the end-to-end framework, none of them have taken into account the characteristic error distribution made by the model. In this paper, we propose a novel approach to utilizing text-only data, by training a spelling correction (SC) model to explicitly correct those errors. On the LibriSpeech dataset, we demonstrate that the proposed model results in an 18.6% relative improvement in WER over the baseline model when directly correcting top ASR hypothesis, and a 29.0% relative improvement when further rescoring an expanded n-best list using an external LM.

This paper presents a novel deep neural network (DNN) architecture with highway blocks (HWs) using a complex discrete Fourier transform (DFT) feature for keyword spotting. In our previous work, we showed that the feed-forward DNN with a time-delayed bottleneck layer (TDB-DNN) directly trained from the audio input outperformed the model with the log-mel filter bank energy feature (LFBE), given a large amount of training data [1]. However, the deeper structure of such an audio input DNN makes an optimization problem more difficult, which could easily fall in one of the local minimum solutions. In order to alleviate the problem, we propose a new HW network with a time-delayed bottleneck layer (TDB-HW). Our TDB-HW networks can learn a bottleneck feature representation through optimization based on the cross-entropy criterion without stage-wise training proposed in [1]. Moreover, we use the complex DFT feature as a method of pre-processing. Our experimental results on the real data show that the TDB-HW network with the complex DFT feature provides significantly lower miss rates for a range of false alarm rates over the LFBE DNN, yielding approximately 20 % relative improvement in the area under the curve (AUC) of the detection error tradeoff (DET) curves for keyword spotting. Furthermore, we investigate the effects of different pre-processing methods for the deep highway network.

This paper proposes an age-dependent scheme for automatic height estimation and speaker normalization of children’s speech, using the first three subglottal resonances (SGRs). Similar to previous work, our analysis indicates that children above the age of 11 years show different acoustic properties from those under 11. Therefore, an age-dependent model is investigated. The estimation algorithms for the first three SGRs are motivated by our previous research for adults. The algorithms for the first two SGRs have been applied to children’s speech before. This paper proposes a similar approach to estimate Sg3 for children. The algorithm is trained and evaluated on 46 children, aged between 6-17 years, using cross-validation. Average RMS errors in estimating Sg1, Sg2 and Sg3 using the age-dependent model are 51, 128 and 168 Hz, respectively. The height estimation algorithm employs a negative correlation between SGRs and height, and the mean absolute height estimation error was found to be less than 3.8cm for the younger children and 4.9cm for the older children. In addition, using TIDIGITS, a linear frequency warping scheme using age-dependent Sg3 gives statisticallysignificant word error rate reductions (up to 26%) relative to conventional VTLN.

The relationship between vowel formants and the second subglottal resonance (Sg2) has previously been explored in English, German, Hungarian and Korean. Results from these studies indicate that vowel space is categorically divided by Sg2 and that Sg2 correlates well with standing height. One of the goals of this work is to verify if the above findings hold true in Mandarin as well. The correlation between Sg2 and sitting height (trunk length) is also studied. Further, since Mandarin is a tonal language (with more pitch variations compared to English), we study the relationship between Sg2 and fundamental frequency (F0). A new corpus of simultaneous recordings of speech and subglottal acoustics was collected from 20 native Mandarin speakers. Results on this corpus indicate that Sg2 divides vowel space in Mandarin as well, and that it is more correlated with sitting height than standing height. Paired t-tests are conducted on the Sg2 measurements from different vowel parts, which represent different F0 regions. Preliminary results show that there is no statistically-significant variation of Sg2 with F0 within a tone. Index Terms: second subglottal resonance, Mandarin, vowel space, sitting height, tonal language