We summarize the results of a host of efforts using giant automatic speech recognition (ASR) models pre-trained using large, diverse unlabeled datasets containing approximately a million hours of audio. We find that the combination of pre-training, self-training and scaling up model size greatly increases data efficiency, even for extremely large tasks with tens of thousands of hours of labeled data. In particular, on an ASR task with 34k hours of labeled data, by fine-tuning an 8 billion parameter pre-trained Conformer model we can match state-of-the-art (SoTA) performance with only 3% of the training data and significantly improve SoTA with the full training set. We also report on the universal benefits gained from using big pre-trained and self-trained models for a large set of downstream tasks that cover a wide range of speech domains and span multiple orders of magnitudes of dataset sizes, including obtaining SoTA performance on many public benchmarks. In addition, we utilize the learned representation of pre-trained networks to achieve SoTA results on non-ASR tasks.
This paper introduces WaveGrad 2, a non-autoregressive generative model for text-to-speech synthesis. WaveGrad 2 is trained to estimate the gradient of the log conditional density of the waveform given a phoneme sequence. The model takes an input phoneme sequence, and through an iterative refinement process, generates an audio waveform. This contrasts to the original WaveGrad vocoder which conditions on mel-spectrogram features, generated by a separate model. The iterative refinement process starts from Gaussian noise, and through a series of refinement steps (e.g., 50 steps), progressively recovers the audio sequence. WaveGrad 2 offers a natural way to trade-off between inference speed and sample quality, through adjusting the number of refinement steps. Experiments show that the model can generate high fidelity audio, approaching the performance of a state-of-the-art neural TTS system. We also report various ablation studies over different model configurations. Audio samples are available at https://wavegrad.github.io/v2.
Denoising Diffusion Probabilistic Models (DDPMs) have emerged as a powerful family of generative models that can yield high-fidelity samples and competitive log-likelihoods across a range of domains, including image and speech synthesis. Key advantages of DDPMs include ease of training, in contrast to generative adversarial networks, and speed of generation, in contrast to autoregressive models. However, DDPMs typically require hundreds-to-thousands of steps to generate a high fidelity sample, making them prohibitively expensive for high dimensional problems. Fortunately, DDPMs allow trading generation speed for sample quality through adjusting the number of refinement steps as a post process. Prior work has been successful in improving generation speed through handcrafting the time schedule by trial and error. We instead view the selection of the inference time schedules as an optimization problem, and introduce an exact dynamic programming algorithm that finds the optimal discrete time schedules for any pre-trained DDPM. Our method exploits the fact that ELBO can be decomposed into separate KL terms, and given any computation budget, discovers the time schedule that maximizes the training ELBO exactly. Our method is efficient, has no hyper-parameters of its own, and can be applied to any pre-trained DDPM with no retraining. We discover inference time schedules requiring as few as 32 refinement steps, while sacrificing less than 0.1 bits per dimension compared to the default 4,000 steps used on ImageNet 64x64 [Ho et al., 2020; Nichol and Dhariwal, 2021].
We present SpeechStew, a speech recognition model that is trained on a combination of various publicly available speech recognition datasets: AMI, Broadcast News, Common Voice, LibriSpeech, Switchboard/Fisher, Tedlium, and Wall Street Journal. SpeechStew simply mixes all of these datasets together, without any special re-weighting or re-balancing of the datasets. SpeechStew achieves SoTA or near SoTA results across a variety of tasks, without the use of an external language model. Our results include 9.0\% WER on AMI-IHM, 4.7\% WER on Switchboard, 8.3\% WER on CallHome, and 1.3\% on WSJ, which significantly outperforms prior work with strong external language models. We also demonstrate that SpeechStew learns powerful transfer learning representations. We fine-tune SpeechStew on a noisy low resource speech dataset, CHiME-6. We achieve 38.9\% WER without a language model, which compares to 38.6\% WER to a strong HMM baseline with a language model.
We present SR3, an approach to image Super-Resolution via Repeated Refinement. SR3 adapts denoising diffusion probabilistic models to conditional image generation and performs super-resolution through a stochastic denoising process. Inference starts with pure Gaussian noise and iteratively refines the noisy output using a U-Net model trained on denoising at various noise levels. SR3 exhibits strong performance on super-resolution tasks at different magnification factors, on faces and natural images. We conduct human evaluation on a standard 8X face super-resolution task on CelebA-HQ, comparing with SOTA GAN methods. SR3 achieves a fool rate close to 50%, suggesting photo-realistic outputs, while GANs do not exceed a fool rate of 34%. We further show the effectiveness of SR3 in cascaded image generation, where generative models are chained with super-resolution models, yielding a competitive FID score of 11.3 on ImageNet.
We combine recent advancements in end-to-end speech recognition to non-autoregressive automatic speech recognition. We push the limits of non-autoregressive state-of-the-art results for multiple datasets: LibriSpeech, Fisher+Switchboard and Wall Street Journal. Key to our recipe, we leverage CTC on giant Conformer neural network architectures with SpecAugment and wav2vec2 pre-training. We achieve 1.8%/3.6% WER on LibriSpeech test/test-other sets, 5.1%/9.8% WER on Switchboard, and 3.4% on the Wall Street Journal, all without a language model.
A channel corresponds to a viewpoint or transformation of an underlying meaning. A pair of parallel sentences in English and French express the same underlying meaning, but through two separate channels corresponding to their languages. In this work, we present the Multichannel Generative Language Model (MGLM). MGLM is a generative joint distribution model over channels. MGLM marginalizes over all possible factorizations within and across all channels. MGLM endows flexible inference, including unconditional generation, conditional generation (where 1 channel is observed and other channels are generated), and partially observed generation (where incomplete observations are spread across all the channels). We experiment with the Multi30K dataset containing English, French, Czech, and German. We demonstrate experiments with unconditional, conditional, and partially conditional generation. We provide qualitative samples sampled unconditionally from the generative joint distribution. We also quantitatively analyze the quality-diversity trade-offs and find MGLM outperforms traditional bilingual discriminative models.
This paper introduces WaveGrad, a conditional model for waveform generation through estimating gradients of the data density. This model is built on the prior work on score matching and diffusion probabilistic models. It starts from Gaussian white noise and iteratively refines the signal via a gradient-based sampler conditioned on the mel-spectrogram. WaveGrad is non-autoregressive, and requires only a constant number of generation steps during inference. It can use as few as 6 iterations to generate high fidelity audio samples. WaveGrad is simple to train, and implicitly optimizes for the weighted variational lower-bound of the log-likelihood. Empirical experiments reveal WaveGrad to generate high fidelity audio samples matching a strong likelihood-based autoregressive baseline with less sequential operations.