Automatic speech recognition (ASR) models are typically designed to operate on a single input data type, e.g. a single or multi-channel audio streamed from a device. This design decision assumes the primary input data source does not change and if an additional (auxiliary) data source is occasionally available, it cannot be used. An ASR model that operates on both primary and auxiliary data can achieve better accuracy compared to a primary-only solution; and a model that can serve both primary-only (PO) and primary-plus-auxiliary (PPA) modes is highly desirable. In this work, we propose a unified ASR model that can serve both modes. We demonstrate its efficacy in a realistic scenario where a set of devices typically stream a single primary audio channel, and two additional auxiliary channels only when upload bandwidth allows it. The architecture enables a unique methodology that uses both types of input audio during training time. Our proposed approach achieves up to 12.5% relative word-error-rate reduction (WERR) compared to a PO baseline, and up to 16.0% relative WERR in low-SNR conditions. The unique training methodology achieves up to 2.5% relative WERR compared to a PPA baseline.
End-to-end (E2E) automatic speech recognition (ASR) models have recently demonstrated superior performance over the traditional hybrid ASR models. Training an E2E ASR model requires a large amount of data which is not only expensive but may also raise dependency on production data. At the same time, synthetic speech generated by the state-of-the-art text-to-speech (TTS) engines has advanced to near-human naturalness. In this work, we propose to utilize synthetic speech for ASR training (SynthASR) in applications where data is sparse or hard to get for ASR model training. In addition, we apply continual learning with a novel multi-stage training strategy to address catastrophic forgetting, achieved by a mix of weighted multi-style training, data augmentation, encoder freezing, and parameter regularization. In our experiments conducted on in-house datasets for a new application of recognizing medication names, training ASR RNN-T models with synthetic audio via the proposed multi-stage training improved the recognition performance on new application by more than 65% relative, without degradation on existing general applications. Our observations show that SynthASR holds great promise in training the state-of-the-art large-scale E2E ASR models for new applications while reducing the costs and dependency on production data.
We propose a simple yet effective method to compress an RNN-Transducer (RNN-T) through the well-known knowledge distillation paradigm. We show that the transducer's encoder outputs naturally have a high entropy and contain rich information about acoustically similar word-piece confusions. This rich information is suppressed when combined with the lower entropy decoder outputs to produce the joint network logits. Consequently, we introduce an auxiliary loss to distill the encoder logits from a teacher transducer's encoder, and explore training strategies where this encoder distillation works effectively. We find that tandem training of teacher and student encoders with an inplace encoder distillation outperforms the use of a pre-trained and static teacher transducer. We also report an interesting phenomenon we refer to as implicit distillation, that occurs when the teacher and student encoders share the same decoder. Our experiments show 5.37-8.4% relative word error rate reductions (WERR) on in-house test sets, and 5.05-6.18% relative WERRs on LibriSpeech test sets.
There is a recent trend in machine learning to increase model quality by growing models to sizes previously thought to be unreasonable. Recent work has shown that autoregressive generative models with cross-entropy objective functions exhibit smooth power-law relationships, or scaling laws, that predict model quality from model size, training set size, and the available compute budget. These scaling laws allow one to choose nearly optimal hyper-parameters given constraints on available training data, model parameter count, or training computation budget. In this paper, we demonstrate that acoustic models trained with an auto-predictive coding loss behave as if they are subject to similar scaling laws. We extend previous work to jointly predict loss due to model size, to training set size, and to the inherent "irreducible loss" of the task. We find that the scaling laws accurately match model performance over two orders of magnitude in both model size and training set size, and make predictions about the limits of model performance.
Text-to-speech systems recently achieved almost indistinguishable quality from human speech. However, the prosody of those systems is generally flatter than natural speech, producing samples with low expressiveness. Disentanglement of speaker id and prosody is crucial in text-to-speech systems to improve on naturalness and produce more variable syntheses. This paper proposes a new neural text-to-speech model that approaches the disentanglement problem by conditioning a Tacotron2-like architecture on flow-normalized speaker embeddings, and by substituting the reference encoder with a new learned latent distribution responsible for modeling the intra-sentence variability due to the prosody. By removing the reference encoder dependency, the speaker-leakage problem typically happening in this kind of systems disappears, producing more distinctive syntheses at inference time. The new model achieves significantly higher prosody variance than the baseline in a set of quantitative prosody features, as well as higher speaker distinctiveness, without decreasing the speaker intelligibility. Finally, we observe that the normalized speaker embeddings enable much richer speaker interpolations, substantially improving the distinctiveness of the new interpolated speakers.
Attention-based beamformers have recently been shown to be effective for multi-channel speech recognition. However, they are less capable at capturing local information. In this work, we propose a 2D Conv-Attention module which combines convolution neural networks with attention for beamforming. We apply self- and cross-attention to explicitly model the correlations within and between the input channels. The end-to-end 2D Conv-Attention model is compared with a multi-head self-attention and superdirective-based neural beamformers. We train and evaluate on an in-house multi-channel dataset. The results show a relative improvement of 3.8% in WER by the proposed model over the baseline neural beamformer.
Comprehending the overall intent of an utterance helps a listener recognize the individual words spoken. Inspired by this fact, we perform a novel study of the impact of explicitly incorporating intent representations as additional information to improve a recurrent neural network-transducer (RNN-T) based automatic speech recognition (ASR) system. An audio-to-intent (A2I) model encodes the intent of the utterance in the form of embeddings or posteriors, and these are used as auxiliary inputs for RNN-T training and inference. Experimenting with a 50k-hour far-field English speech corpus, this study shows that when running the system in non-streaming mode, where intent representation is extracted from the entire utterance and then used to bias streaming RNN-T search from the start, it provides a 5.56% relative word error rate reduction (WERR). On the other hand, a streaming system using per-frame intent posteriors as extra inputs for the RNN-T ASR system yields a 3.33% relative WERR. A further detailed analysis of the streaming system indicates that our proposed method brings especially good gain on media-playing related intents (e.g. 9.12% relative WERR on PlayMusicIntent).
Wav2vec-C introduces a novel representation learning technique combining elements from wav2vec 2.0 and VQ-VAE. Our model learns to reproduce quantized representations from partially masked speech encoding using a contrastive loss in a way similar to Wav2vec 2.0. However, the quantization process is regularized by an additional consistency network that learns to reconstruct the input features to the wav2vec 2.0 network from the quantized representations in a way similar to a VQ-VAE model. The proposed self-supervised model is trained on 10k hours of unlabeled data and subsequently used as the speech encoder in a RNN-T ASR model and fine-tuned with 1k hours of labeled data. This work is one of only a few studies of self-supervised learning on speech tasks with a large volume of real far-field labeled data. The Wav2vec-C encoded representations achieves, on average, twice the error reduction over baseline and a higher codebook utilization in comparison to wav2vec 2.0
Spoken language understanding (SLU) systems extract transcriptions, as well as semantics of intent or named entities from speech, and are essential components of voice activated systems. SLU models, which either directly extract semantics from audio or are composed of pipelined automatic speech recognition (ASR) and natural language understanding (NLU) models, are typically trained via differentiable cross-entropy losses, even when the relevant performance metrics of interest are word or semantic error rates. In this work, we propose non-differentiable sequence losses based on SLU metrics as a proxy for semantic error and use the REINFORCE trick to train ASR and SLU models with this loss. We show that custom sequence loss training is the state-of-the-art on open SLU datasets and leads to 6% relative improvement in both ASR and NLU performance metrics on large proprietary datasets. We also demonstrate how the semantic sequence loss training paradigm can be used to update ASR and SLU models without transcripts, using semantic feedback alone.