Michiel Bacchiani

Yuma Koizumi, Heiga Zen, Shigeki Karita et al.

Speech restoration (SR) is a task of converting degraded speech signals into high-quality ones. In this study, we propose a robust SR model called Miipher, and apply Miipher to a new SR application: increasing the amount of high-quality training data for speech generation by converting speech samples collected from the Web to studio-quality. To make our SR model robust against various degradation, we use (i) a speech representation extracted from w2v-BERT for the input feature, and (ii) a text representation extracted from transcripts via PnG-BERT as a linguistic conditioning feature. Experiments show that Miipher (i) is robust against various audio degradation and (ii) enable us to train a high-quality text-to-speech (TTS) model from restored speech samples collected from the Web. Audio samples are available at our demo page: google.github.io/df-conformer/miipher/

Yuma Koizumi, Kohei Yatabe, Heiga Zen et al.

Denoising diffusion probabilistic models (DDPMs) and generative adversarial networks (GANs) are popular generative models for neural vocoders. The DDPMs and GANs can be characterized by the iterative denoising framework and adversarial training, respectively. This study proposes a fast and high-quality neural vocoder called \textit{WaveFit}, which integrates the essence of GANs into a DDPM-like iterative framework based on fixed-point iteration. WaveFit iteratively denoises an input signal, and trains a deep neural network (DNN) for minimizing an adversarial loss calculated from intermediate outputs at all iterations. Subjective (side-by-side) listening tests showed no statistically significant differences in naturalness between human natural speech and those synthesized by WaveFit with five iterations. Furthermore, the inference speed of WaveFit was more than 240 times faster than WaveRNN. Audio demos are available at \url{google.github.io/df-conformer/wavefit/}.

Min Ma, Yuma Koizumi, Shigeki Karita et al.

This paper introduces FLEURS-R, a speech restoration applied version of the Few-shot Learning Evaluation of Universal Representations of Speech (FLEURS) corpus. FLEURS-R maintains an N-way parallel speech corpus in 102 languages as FLEURS, with improved audio quality and fidelity by applying the speech restoration model Miipher. The aim of FLEURS-R is to advance speech technology in more languages and catalyze research including text-to-speech (TTS) and other speech generation tasks in low-resource languages. Comprehensive evaluations with the restored speech and TTS baseline models trained from the new corpus show that the new corpus obtained significantly improved speech quality while maintaining the semantic contents of the speech. The corpus is publicly released via Hugging Face.

Chanwoo Kim, Ehsan Variani, Arun Narayanan et al.

In this paper, we describe how to efficiently implement an acoustic room simulator to generate large-scale simulated data for training deep neural networks. Even though Google Room Simulator in [1] was shown to be quite effective in reducing the Word Error Rates (WERs) for far-field applications by generating simulated far-field training sets, it requires a very large number of Fast Fourier Transforms (FFTs) of large size. Room Simulator in [1] used approximately 80 percent of Central Processing Unit (CPU) usage in our CPU + Graphics Processing Unit (GPU) training architecture [2]. In this work, we implement an efficient OverLap Addition (OLA) based filtering using the open-source FFTW3 library. Further, we investigate the effects of the Room Impulse Response (RIR) lengths. Experimentally, we conclude that we can cut the tail portions of RIRs whose power is less than 20 dB below the maximum power without sacrificing the speech recognition accuracy. However, we observe that cutting RIR tail more than this threshold harms the speech recognition accuracy for rerecorded test sets. Using these approaches, we were able to reduce CPU usage for the room simulator portion down to 9.69 percent in CPU/GPU training architecture. Profiling result shows that we obtain 22.4 times speed-up on a single machine and 37.3 times speed up on Google's distributed training infrastructure.

Shigeki Karita, Yuma Koizumi, Heiga Zen et al.

Training data cleaning is a new application for generative model-based speech restoration (SR). This paper introduces Miipher-2, an SR model designed for million-hour scale data, for training data cleaning for large-scale generative models like large language models. Key challenges addressed include generalization to unseen languages, operation without explicit conditioning (e.g., text, speaker ID), and computational efficiency. Miipher-2 utilizes a frozen, pre-trained Universal Speech Model (USM), supporting over 300 languages, as a robust, conditioning-free feature extractor. To optimize efficiency and minimize memory, Miipher-2 incorporates parallel adapters for predicting clean USM features from noisy inputs and employs the WaveFit neural vocoder for waveform synthesis. These components were trained on 3,000 hours of multi-lingual, studio-quality recordings with augmented degradations, while USM parameters remained fixed. Experimental results demonstrate Miipher-2's superior or comparable performance to conventional SR models in word-error-rate, speaker similarity, and both objective and subjective sound quality scores across all tested languages. Miipher-2 operates efficiently on consumer-grade accelerators, achieving a real-time factor of 0.0078, enabling the processing of a million-hour speech dataset in approximately three days using only 100 such accelerators.

Yotaro Kubo, Xingyu Cai, Michiel Bacchiani

This paper proposes a method to effectively perform joint training-and-pruning based on adaptive dropout layers with unit-wise retention probabilities. The proposed method is based on the estimation of a unit-wise retention probability in a dropout layer. A unit that is estimated to have a small retention probability can be considered to be prunable. The retention probability of the unit is estimated using back-propagation and the Gumbel-Softmax technique. This pruning method is applied at several application points in Conformers such that the effective number of parameters can be significantly reduced. Specifically, adaptive dropout layers are introduced in three locations in each Conformer block: (a) the hidden layer of the feed-forward-net component, (b) the query vectors and the value vectors of the self-attention component, and (c) the input vectors of the LConv component. The proposed method is evaluated by conducting a speech recognition experiment on the LibriSpeech task. It was shown that this approach could simultaneously achieve a parameter reduction and accuracy improvement. The word error rates improved by approx 1% while reducing the number of parameters by 54%.

Yuma Koizumi, Heiga Zen, Kohei Yatabe et al.

Neural vocoder using denoising diffusion probabilistic model (DDPM) has been improved by adaptation of the diffusion noise distribution to given acoustic features. In this study, we propose SpecGrad that adapts the diffusion noise so that its time-varying spectral envelope becomes close to the conditioning log-mel spectrogram. This adaptation by time-varying filtering improves the sound quality especially in the high-frequency bands. It is processed in the time-frequency domain to keep the computational cost almost the same as the conventional DDPM-based neural vocoders. Experimental results showed that SpecGrad generates higher-fidelity speech waveform than conventional DDPM-based neural vocoders in both analysis-synthesis and speech enhancement scenarios. Audio demos are available at wavegrad.github.io/specgrad/.

Yotaro Kubo, Shigeki Karita, Michiel Bacchiani

End-to-end speech recognition is a promising technology for enabling compact automatic speech recognition (ASR) systems since it can unify the acoustic and language model into a single neural network. However, as a drawback, training of end-to-end speech recognizers always requires transcribed utterances. Since end-to-end models are also known to be severely data hungry, this constraint is crucial especially because obtaining transcribed utterances is costly and can possibly be impractical or impossible. This paper proposes a method for alleviating this issue by transferring knowledge from a language model neural network that can be pretrained with text-only data. Specifically, this paper attempts to transfer semantic knowledge acquired in embedding vectors of large-scale language models. Since embedding vectors can be assumed as implicit representations of linguistic information such as part-of-speech, intent, and so on, those are also expected to be useful modeling cues for ASR decoders. This paper extends two types of ASR decoders, attention-based decoders and neural transducers, by modifying training loss functions to include embedding prediction terms. The proposed systems were shown to be effective for error rate reduction without incurring extra computational costs in the decoding phase.

Yuma Koizumi, Shigeki Karita, Arun Narayanan et al.

Speech enhancement (SE) is used as a frontend in speech applications including automatic speech recognition (ASR) and telecommunication. A difficulty in using the SE frontend is that the appropriate noise reduction level differs depending on applications and/or noise characteristics. In this study, we propose "signal-to-noise ratio improvement (SNRi) target training"; the SE frontend is trained to output a signal whose SNRi is controlled by an auxiliary scalar input. In joint training with a backend, the target SNRi value is estimated by an auxiliary network. By training all networks to minimize the backend task loss, we can estimate the appropriate noise reduction level for each noisy input in a data-driven scheme. Our experiments showed that the SNRi target training enables control of the output SNRi. In addition, the proposed joint training relatively reduces word error rate by 4.0\% and 5.7\% compared to a Conformer-based standard ASR model and conventional SE-ASR joint training model, respectively. Furthermore, by analyzing the predicted target SNRi, we observed the jointly trained network automatically controls the target SNRi according to noise characteristics. Audio demos are available in our demo page: google.github.io/df-conformer/snri_target/.

Yuma Koizumi, Shigeki Karita, Scott Wisdom et al.

Single-channel speech enhancement (SE) is an important task in speech processing. A widely used framework combines an analysis/synthesis filterbank with a mask prediction network, such as the Conv-TasNet architecture. In such systems, the denoising performance and computational efficiency are mainly affected by the structure of the mask prediction network. In this study, we aim to improve the sequential modeling ability of Conv-TasNet architectures by integrating Conformer layers into a new mask prediction network. To make the model computationally feasible, we extend the Conformer using linear complexity attention and stacked 1-D dilated depthwise convolution layers. We trained the model on 3,396 hours of noisy speech data, and show that (i) the use of linear complexity attention avoids high computational complexity, and (ii) our model achieves higher scale-invariant signal-to-noise ratio than the improved time-dilated convolution network (TDCN++), an extended version of Conv-TasNet.

Parisa Haghani, Arun Narayanan, Michiel Bacchiani et al.

Conventional spoken language understanding systems consist of two main components: an automatic speech recognition module that converts audio to a transcript, and a natural language understanding module that transforms the resulting text (or top N hypotheses) into a set of domains, intents, and arguments. These modules are typically optimized independently. In this paper, we formulate audio to semantic understanding as a sequence-to-sequence problem [1]. We propose and compare various encoder-decoder based approaches that optimize both modules jointly, in an end-to-end manner. Evaluations on a real-world task show that 1) having an intermediate text representation is crucial for the quality of the predicted semantics, especially the intent arguments and 2) jointly optimizing the full system improves overall accuracy of prediction. Compared to independently trained models, our best jointly trained model achieves similar domain and intent prediction F1 scores, but improves argument word error rate by 18% relative.

Arun Narayanan, Ananya Misra, Khe Chai Sim et al.

Current state-of-the-art automatic speech recognition systems are trained to work in specific `domains', defined based on factors like application, sampling rate and codec. When such recognizers are used in conditions that do not match the training domain, performance significantly drops. This work explores the idea of building a single domain-invariant model for varied use-cases by combining large scale training data from multiple application domains. Our final system is trained using 162,000 hours of speech. Additionally, each utterance is artificially distorted during training to simulate effects like background noise, codec distortion, and sampling rates. Our results show that, even at such a scale, a model thus trained works almost as well as those fine-tuned to specific subsets: A single model can be robust to multiple application domains, and variations like codecs and noise. More importantly, such models generalize better to unseen conditions and allow for rapid adaptation -- we show that by using as little as 10 hours of data from a new domain, an adapted domain-invariant model can match performance of a domain-specific model trained from scratch using 70 times as much data. We also highlight some of the limitations of such models and areas that need addressing in future work.

Chung-Cheng Chiu, Tara N. Sainath, Yonghui Wu et al.

Attention-based encoder-decoder architectures such as Listen, Attend, and Spell (LAS), subsume the acoustic, pronunciation and language model components of a traditional automatic speech recognition (ASR) system into a single neural network. In previous work, we have shown that such architectures are comparable to state-of-theart ASR systems on dictation tasks, but it was not clear if such architectures would be practical for more challenging tasks such as voice search. In this work, we explore a variety of structural and optimization improvements to our LAS model which significantly improve performance. On the structural side, we show that word piece models can be used instead of graphemes. We also introduce a multi-head attention architecture, which offers improvements over the commonly-used single-head attention. On the optimization side, we explore synchronous training, scheduled sampling, label smoothing, and minimum word error rate optimization, which are all shown to improve accuracy. We present results with a unidirectional LSTM encoder for streaming recognition. On a 12, 500 hour voice search task, we find that the proposed changes improve the WER from 9.2% to 5.6%, while the best conventional system achieves 6.7%; on a dictation task our model achieves a WER of 4.1% compared to 5% for the conventional system.

Bo Li, Tara N. Sainath, Khe Chai Sim et al.

Sequence-to-sequence models provide a simple and elegant solution for building speech recognition systems by folding separate components of a typical system, namely acoustic (AM), pronunciation (PM) and language (LM) models into a single neural network. In this work, we look at one such sequence-to-sequence model, namely listen, attend and spell (LAS), and explore the possibility of training a single model to serve different English dialects, which simplifies the process of training multi-dialect systems without the need for separate AM, PM and LMs for each dialect. We show that simply pooling the data from all dialects into one LAS model falls behind the performance of a model fine-tuned on each dialect. We then look at incorporating dialect-specific information into the model, both by modifying the training targets by inserting the dialect symbol at the end of the original grapheme sequence and also feeding a 1-hot representation of the dialect information into all layers of the model. Experimental results on seven English dialects show that our proposed system is effective in modeling dialect variations within a single LAS model, outperforming a LAS model trained individually on each of the seven dialects by 3.1 ~ 16.5% relative.