Synthetic data generated by text-to-speech (TTS) systems can be used to improve automatic speech recognition (ASR) systems in low-resource or domain mismatch tasks. It has been shown that TTS-generated outputs still do not have the same qualities as real data. In this work we focus on the temporal structure of synthetic data and its relation to ASR training. By using a novel oracle setup we show how much the degradation of synthetic data quality is influenced by duration modeling in non-autoregressive (NAR) TTS. To get reference phoneme durations we use two common alignment methods, a hidden Markov Gaussian-mixture model (HMM-GMM) aligner and a neural connectionist temporal classification (CTC) aligner. Using a simple algorithm based on random walks we shift phoneme duration distributions of the TTS system closer to real durations, resulting in an improvement of an ASR system using synthetic data in a semi-supervised setting.
In this work, we investigate the effect of language models (LMs) with different context lengths and label units (phoneme vs. word) used in sequence discriminative training for phoneme-based neural transducers. Both lattice-free and N-best-list approaches are examined. For lattice-free methods with phoneme-level LMs, we propose a method to approximate the context history to employ LMs with full-context dependency. This approximation can be extended to arbitrary context length and enables the usage of word-level LMs in lattice-free methods. Moreover, a systematic comparison is conducted across lattice-free and N-best-list-based methods. Experimental results on Librispeech show that using the word-level LM in training outperforms the phoneme-level LM. Besides, we find that the context size of the LM used for probability computation has a limited effect on performance. Moreover, our results reveal the pivotal importance of the hypothesis space quality in sequence discriminative training.
We investigate a novel modeling approach for end-to-end neural network training using hidden Markov models (HMM) where the transition probabilities between hidden states are modeled and learned explicitly. Most contemporary sequence-to-sequence models allow for from-scratch training by summing over all possible label segmentations in a given topology. In our approach there are explicit, learnable probabilities for transitions between segments as opposed to a blank label that implicitly encodes duration statistics. We implement a GPU-based forward-backward algorithm that enables the simultaneous training of label and transition probabilities. We investigate recognition results and additionally Viterbi alignments of our models. We find that while the transition model training does not improve recognition performance, it has a positive impact on the alignment quality. The generated alignments are shown to be viable targets in state-of-the-art Viterbi trainings.
Internal language model (ILM) subtraction has been widely applied to improve the performance of the RNN-Transducer with external language model (LM) fusion for speech recognition. In this work, we show that sequence discriminative training has a strong correlation with ILM subtraction from both theoretical and empirical points of view. Theoretically, we derive that the global optimum of maximum mutual information (MMI) training shares a similar formula as ILM subtraction. Empirically, we show that ILM subtraction and sequence discriminative training achieve similar performance across a wide range of experiments on Librispeech, including both MMI and minimum Bayes risk (MBR) criteria, as well as neural transducers and LMs of both full and limited context. The benefit of ILM subtraction also becomes much smaller after sequence discriminative training. We also provide an in-depth study to show that sequence discriminative training has a minimal effect on the commonly used zero-encoder ILM estimation, but a joint effect on both encoder and prediction + joint network for posterior probability reshaping including both ILM and blank suppression.
Many real-life applications of automatic speech recognition (ASR) require processing of overlapped speech. A commonmethod involves first separating the speech into overlap-free streams and then performing ASR on the resulting signals. Recently, the inclusion of a mixture encoder in the ASR model has been proposed. This mixture encoder leverages the original overlapped speech to mitigate the effect of artifacts introduced by the speech separation. Previously, however, the method only addressed two-speaker scenarios. In this work, we extend this approach to more natural meeting contexts featuring an arbitrary number of speakers and dynamic overlaps. We evaluate the performance using different speech separators, including the powerful TF-GridNet model. Our experiments show state-of-the-art performance on the LibriCSS dataset and highlight the advantages of the mixture encoder. Furthermore, they demonstrate the strong separation of TF-GridNet which largely closes the gap between previous methods and oracle separation.
We study a streamable attention-based encoder-decoder model in which either the decoder, or both the encoder and decoder, operate on pre-defined, fixed-size windows called chunks. A special end-of-chunk (EOC) symbol advances from one chunk to the next chunk, effectively replacing the conventional end-of-sequence symbol. This modification, while minor, situates our model as equivalent to a transducer model that operates on chunks instead of frames, where EOC corresponds to the blank symbol. We further explore the remaining differences between a standard transducer and our model. Additionally, we examine relevant aspects such as long-form speech generalization, beam size, and length normalization. Through experiments on Librispeech and TED-LIUM-v2, and by concatenating consecutive sequences for long-form trials, we find that our streamable model maintains competitive performance compared to the non-streamable variant and generalizes very well to long-form speech.
Automatic speech recognition (ASR) systems typically use handcrafted feature extraction pipelines. To avoid their inherent information loss and to achieve more consistent modeling from speech to transcribed text, neural raw waveform feature extractors (FEs) are an appealing approach. Also the wav2vec 2.0 model, which has recently gained large popularity, uses a convolutional FE which operates directly on the speech waveform. However, it is not yet studied extensively in the literature. In this work, we study its capability to replace the standard feature extraction methods in a connectionist temporal classification (CTC) ASR model and compare it to an alternative neural FE. We show that both are competitive with traditional FEs on the LibriSpeech benchmark and analyze the effect of the individual components. Furthermore, we analyze the learned filters and show that the most important information for the ASR system is obtained by a set of bandpass filters.
Multi-speaker automatic speech recognition (ASR) is crucial for many real-world applications, but it requires dedicated modeling techniques. Existing approaches can be divided into modular and end-to-end methods. Modular approaches separate speakers and recognize each of them with a single-speaker ASR system. End-to-end models process overlapped speech directly in a single, powerful neural network. This work proposes a middle-ground approach that leverages explicit speech separation similarly to the modular approach but also incorporates mixture speech information directly into the ASR module in order to mitigate the propagation of errors made by the speech separator. We also explore a way to exchange cross-speaker context information through a layer that combines information of the individual speakers. Our system is optimized through separate and joint training stages and achieves a relative improvement of 7% in word error rate over a purely modular setup on the SMS-WSJ task.
Building competitive hybrid hidden Markov model~(HMM) systems for automatic speech recognition~(ASR) requires a complex multi-stage pipeline consisting of several training criteria. The recent sequence-to-sequence models offer the advantage of having simpler pipelines that can start from-scratch. We propose a purely neural based single-stage from-scratch pipeline for a context-dependent hybrid HMM that offers similar simplicity. We use an alignment from a full-sum trained zero-order posterior HMM with a BLSTM encoder. We show that with this alignment we can build a Conformer factored hybrid that performs even better than both a state-of-the-art classic hybrid and a factored hybrid trained with alignments taken from more complex Gaussian mixture based systems. Our finding is confirmed on Switchboard 300h and LibriSpeech 960h tasks with comparable results to other approaches in the literature, and by additionally relying on a responsible choice of available computational resources.