Self-training in the teacher-student framework generally suffers from the confirmation bias problem, where errors from the teacher are propagated to the student and hence get amplified with multiple iterations. In this paper, we present Monte Carlo Self-training where pseudo labels are generated by sampling from a teacher distribution, as a way to mitigate this problem. We show that Monte Carlo Self-training is an approximation to minimizing label sequence level cross entropy between student and teacher. In our experiments we find that Monte Carlo Self-training always outperforms beam decoder based self-training and is quite robust even when the initial teacher WER is high. We also show that label sampling allows formulating pseudo label confidence in a more natural way and we show that these confidence measures give further improvements in our unsupervised adaptation experiments especially when the initial teacher WER is very high.

We introduce a multilingual speaker change detection model (USM-SCD) that can simultaneously detect speaker turns and perform ASR for 96 languages. This model is adapted from a speech foundation model trained on a large quantity of supervised and unsupervised data, demonstrating the utility of fine-tuning from a large generic foundation model for a downstream task. We analyze the performance of this multilingual speaker change detection model through a series of ablation studies. We show that the USM-SCD model can achieve more than 75% average speaker change detection F1 score across a test set that consists of data from 96 languages. On American English, the USM-SCD model can achieve an 85.8% speaker change detection F1 score across various public and internal test sets, beating the previous monolingual baseline model by 21% relative. We also show that we only need to fine-tune one-quarter of the trainable model parameters to achieve the best model performance. The USM-SCD model exhibits state-of-the-art ASR quality compared with a strong public ASR baseline, making it suitable to handle both tasks with negligible additional computational cost.

Semi-supervised training can be performed by jointly optimizing supervised and unsupervised losses. In many settings, supervised and unsupervised losses are inconsistent, and this inconsistency creates instability in training. As a solution, we propose cross-training: instead of training one network with two losses, we train two separate networks, each with a different loss; we then tie the parameters of the networks by minimizing an additional L2 loss between the parameters. This L2 loss acts as a knowledge bridge between the networks. It forces the networks to be similar; therefore both can learn from each other. This paper introduces the cross-training scheme to develop a stable contrastive siamese (c-siam) network. Our experiments on LibriSpeech and Google’s Voice-Search/YouTube datasets show that (1) cross-training provides 20% relative WER improvement over the SOTA systems on the LibriSpeech dataset; (2) cross-training stabilizes c-siam training and significantly outperforms SOTA systems on small supervised datasets; (3) cross-training is effective for cascaded encoders, unlike the original c-siam which shows weak convergence characteristics.

In this work we propose a novel token-based training strategy that improves Transformer-Transducer (T-T) based speaker change detection (SCD) performance. The conventional T-T based SCD model loss optimizes all output tokens equally. Due to the sparsity of the speaker changes in the training data, the conventional T-T based SCD model loss leads to sub-optimal detection accuracy. To mitigate this issue, we use a customized edit-distance algorithm to estimate the token-level SCD false accept (FA) and false reject (FR) rates during training and optimize model parameters to minimize a weighted combination of the FA and FR, focusing the model on accurately predicting speaker changes. We also propose a set of evaluation metrics that align better with commercial use cases. Experiments on a group of challenging real-world datasets show that the proposed training method can significantly improve the overall performance of the SCD model with the same number of parameters.

This paper introduces contrastive siamese (c-siam) network, an architecture for leveraging unlabeled acoustic data in speech recognition. c-siam is the first network that extracts high-level linguistic information from speech by matching outputs of two identical transformer encoders. It contains augmented and target branches which are trained by: (1) masking inputs and matching outputs with a contrastive loss, (2) incorporating a stop gradient operation on the target branch, (3) using an extra learnable transformation on the augmented branch, (4) introducing new temporal augment functions to prevent the shortcut learning problem. We use the Libri-light 60k unsupervised data and the LibriSpeech 100hrs/960hrs supervised data to compare c-siam and other best-performing systems. Our experiments show that c-siam provides 20% relative word error rate improvement over wav2vec baselines. A c-siam network with 450M parameters achieves competitive results compared to the state-of-the-art networks with 600M parameters.

In this paper, we present a novel speaker diarization system for streaming on-device applications. In this system, we use a transformer transducer to detect the speaker turns, represent each speaker turn by a speaker embedding, then cluster these embeddings with constraints from the detected speaker turns. Compared with conventional clustering-based diarization systems, our system largely reduces the computational cost of clustering due to the sparsity of speaker turns. Unlike other supervised speaker diarization systems which require annotations of time-stamped speaker labels for training, our system only requires including speaker turn tokens during the transcribing process, which largely reduces the human efforts involved in data collection.

Reducing prediction delay for streaming end-to-end ASR models with minimal performance regression is a challenging problem. Constrained alignment is a well-known existing approach that penalizes predicted word boundaries using external low-latency acoustic models. On the contrary, recently proposed FastEmit is a sequence-level delay regularization scheme encouraging vocabulary tokens over blanks without any reference alignments. Although all these schemes are successful in reducing delay, ASR word error rate (WER) often severely degrades after applying these delay constraining schemes. In this paper, we propose a novel delay constraining method, named self alignment. Self alignment does not require external alignment models. Instead, it utilizes Viterbi forced-alignments from the trained model to find the lower latency alignment direction. From LibriSpeech evaluation, self alignment outperformed existing schemes: 25% and 56% less delay compared to FastEmit and constrained alignment at the similar word error rate. For Voice Search evaluation,12% and 25% delay reductions were achieved compared to FastEmit and constrained alignment with more than 2% WER improvements.

In this paper we present an end-to-end speech recognition system that can recognize single-channel speech where multiple talkers can speak at the same time (overlapping speech) by using a neural network model based on Recurrent Neural Network Transducer (RNN-T) architecture. We augment the conventional RNN-T architecture by including a masking model for separation of encoded audio features, and multiple label encoders to encode transcripts from different speakers. We use a masking L2 loss to prevent transcripts to align to wrong speakers' audio, and a speaker embedding loss to facilitate speaker tracking. We show that by using these additional training objectives, the proposed augmented RNN-T model can be trained with simulated overlapping speech data and can achieve a WER of 32% on words in overlapping speech segments from real-life telephone conversations. Our analysis of manual transcription task on the same test set shows that transcribing overlapping speech is hard even for humans who can get a WER of 37% compared to ground-truth.

In this paper we present an end-to-end speech recognition model with Transformer encoders that can be used in a streaming speech recognition system. Transformer computation blocks based on self-attention are used to encode both audio and label sequences independently. The activations from both audio and label encoders are combined with a feed-forward layer to compute a probability distribution over the label space for every combination of acoustic frame position and label history. This is similar to the Recurrent Neural Network Transducer (RNN-T) model, which uses RNNs for information encoding instead of Transformer encoders. The model is trained with the RNN-T loss well-suited to streaming decoding. We present results on the LibriSpeech dataset showing that limiting the left context for self-attention in the Transformer layers makes decoding computationally tractable for streaming, with only a slight degradation in accuracy. We also show that the full attention version of our model beats the-state-of-the art accuracy on the LibriSpeech benchmarks. Our results also show that we can bridge the gap between full attention and limited attention versions of our model by attending to a limited number of future frames.

Recurrent Neural Network Transducer (RNNT) is an end-to-end model which transduces discrete input sequences to output sequences by learning alignments between the sequences. In speech recognition tasks we generally have a strictly monotonic alignment between time frames and label sequence. However, the standard RNNT loss does not enforce this constraint. This can cause some anomalies in alignments such as the model outputting a sequence of labels at a single time frame. There is also no bound on the decoding time steps. To address these problems, we introduce a monotonic version of the RNNT loss. Under the assumption that the output sequence is not longer than the input sequence, this loss can be used with forward-backward algorithm to learn strictly monotonic alignments between the sequences. We present experimental studies showing that speech recognition accuracy for monotonic RNNT is equivalent to standard RNNT. We also explore best-first and breadth-first decoding strategies for both monotonic and standard RNNT models. Our experiments show that breadth-first search is effective in exploring and combining alternative alignments. Additionally, it also allows batching of hypotheses during search label expansion, allowing better resource utilization, and resulting in decoding speedup.