Neural networks benefit from large quantities of labeled training data. However, in many settings labeled data is much harder to come by than unlabeled data: current speech recognition systems require thousands of hours of transcribed speech to reach acceptable performance which is not available for the vast majority of the nearly 7,000 languages spoken worldwide (Lewis et al., 2016). Learning purely from labeled examples does not resemble language acquisition in humans: infants learn language by listening to adults around them - a process that requires learning good representations of speech.
In machine learning, self-supervised learning has emerged as a paradigm to learn general data representations from unlabeled examples and to fine-tune the model on labeled data. This has been particularly successful for natural language processing(Peters et al., 2018; Radford et al., 2018; Devlin et al., 2018)
and is an active research area for computer vision(Hénaff et al., 2019; Bachman et al., 2019; Misra and van der Maaten, 2019; He et al., 2019; Chen et al., 2020).
In this paper, we present a framework for self-supervised learning of representations from raw audio data. Our approach encodes speech audio via a multi-layer convolutional neural network and then masks spans of the resulting latent speech representations(Jiang et al., 2019; Wang et al., 2020), similar to masked language modeling (Devlin et al., 2018)
. The latent representations are fed to a Transformer network to build contextualized representations and the model is trained via a contrastive task where the true latent is to be distinguished from distractors(van den Oord et al., 2017; Schneider et al., 2019; Rivière et al., 2020; Kawakami et al., 2020) (§ 2).
As part of training, we learn discrete linguistic units (van den Oord et al., 2017; Liu et al., 2019a; Chorowski et al., 2019; Harwath et al., 2020) via a gumbel softmax (Jang et al., 2016; Baevski et al., 2020) to represent the latent representations in the contrastive task (Figure 1) which we find to be more effective than non-quantized targets. After pre-training on unlabeled speech, the model is fine-tuned on labeled data with a Connectionist Temporal Classification (CTC) loss (Graves et al., 2006; Baevski et al., 2019) to be used for downstream speech recognition tasks (§ 3)
Previous work learned a quantization of the data followed by a contextualized representations with a self-attention model(Baevski et al., 2020, 2019), whereas our approach solves both problems end-to-end. Masking parts of the input with Transformer networks for speech has been explored (Baevski et al., 2019; Jiang et al., 2019), but prior work relies either on a two-step pipeline or their model is trained by reconstructing the filter bank input features. Other related work includes learning representations from auto-encoding the input data (Tjandra et al., 2019; Eloff et al., 2019) or directly predicting future timesteps (Chung et al., 2019).
Our results demonstrate the feasibility of ultra-low resource speech recognition: when using only 10 minutes of labeled data, our approach achieves word error rate (WER) 5.7/10.1 on the clean/noisy test sets of Librispeech. We set a new state of the art on TIMIT phoneme recognition as well as the 100 hour clean subset of Librispeech. Moreover, when we lower the amount of labeled data to just one hour, we still outperform the previous state of the art self-training method of (Park et al., 2020) while using 100 times less labeled data and the same amount of unlabeled data. When we use all 960 hours of labeled data from Librispeech, then our model achieves 1.9/3.5 WER which performs competitively to the best published result while using a simpler baseline architecture (§ 4, § 5).
Our model is composed of a multi-layer convolutional feature encoder which takes as input raw audio and outputs latent speech representations . They are then fed to a Transformer to build representations capturing information from the entire sequence Devlin et al. (2018); Baevski et al. (2020, 2019). The output of the feature encoder is discretized to with a quantization module to represent the targets (Figure 1) in the self-supervised objective (§ 3.2). Compared to vq-wav2vec (Baevski et al., 2020), our model builds context representations over continuous speech representations and self-attention captures dependencies over the entire sequence of latent representations end-to-end.
The encoder consists of several blocks containing a temporal convolution followed by a GELU activation function(Hendrycks and Gimpel, 2016). The first block maps raw audio to a feature representation and to increase robustness, we add a group normalization before the GELU to normalize each output channel over the sequence. We apply layer normalization to the output channels of this network Ba et al. (2016).
Contextualized representations with Transformers.
The output of the feature encoder is fed to a context network which follows the Transformer architecture Vaswani et al. (2017); Devlin et al. (2018); Liu et al. (2019b). Instead of fixed positional embeddings which encode absolute positional information, we use a convolutional layer with kernel size 128 and 16 groups similar to Mohamed et al. (2019); Baevski et al. (2019); Wu et al. (2019) which acts as relative positional embedding. We add the output of the convolution followed by a GELU to the inputs and then apply layer normalization.
For self-supervised training we discretize the output of the feature encoder to a finite set of speech representations via product quantization (Jegou et al., 2011; Baevski et al., 2020). This amounts to choosing quantized representations from multiple codebooks and concatenating them. Given codebooks, or groups, with entries
, we choose one entry from each codebook and concatenate the resulting vectors
and apply a linear transformationto obtain .
. We use the straight-through estimator(Jiang et al., 2019) and setup hard Gumbel softmax operations (Jang et al., 2016). The feature encoder output is mapped to logits and the probabilities for choosing the -th codebook entry for group are
where is a non-negative temperature, and are uniform samples from . During the forward pass, codeword is chosen by and in the backward pass, the true gradient of the Gumbel softmax outputs is used.
To pre-train the model we mask a certain proportion of time steps in the latent feature encoder space (§ 3.1), similar to masked language modeling in BERT (Devlin et al., 2018). The training objective requires identifying the correct quantized latent audio representation in a set of distractors for each masked time step (§ 3.2) and the final model is fine-tuned on the labeled data (§ 3.3).
We mask a proportion of the feature encoder outputs, or time steps before feeding them to the context network and replace them with a trained feature vector shared between all masked time steps; we do not mask inputs to the quantization module. To mask the latent speech representations output by the encoder, we randomly sample without replacement of all time steps to be starting indices and then mask the subsequent consecutive time steps from every sampled index; spans may overlap. This results in approximately 49% of all time steps to be masked with a mean span length of 14.7, or 299ms (see Appendix A for more details on masking) .
During pre-training, we learn representations of speech audio by solving a contrastive task which requires to identify the true quantized latent speech representation for a masked time step within a set of distractors. This is augmented by a codebook diversity loss to encourage the model to use the codebook entries equally often. Finally, we add a L2 penalty over the outputs of the feature encoder:
are tuned hyperparameters.
Given context network output centered over masked time step , the model needs to identify the true quantized latent speech representation in a set of quantized candidate representations which includes and distractors (Hyvärinen, 2010; van den Oord et al., 2018). Distractors are uniformly sampled from other masked time steps of the same utterance. The loss is defined as
where we compute the cosine similaritybetween context representations and quantized latent speech representations (He et al., 2019; Chen et al., 2020).
The contrastive task depends on the codebook to represent both positive and negative examples and the diversity loss is designed to increase the use of the quantized codebook representations (Dieleman et al., 2018). We encourage the equal use of the entries in each of the codebooks by maximizing the entropy of the averaged softmax distribution over the codebook entries for each codebook across a batch of utterances; the softmax disribution does not contain the gumbel noise nor a temperature:222Our implementation maximizes perplexity which is equivalent.
Stabilizing the Feature Encoder.
To stabilize training, we found it helpful to apply an L2 penalty to the activations of the final layer of the feature encoder but before the final layer normalization. We also scale down the global learning for weight updates to the feature encoder by , see § 4.2.
Pre-trained models are fine-tuned for speech recognition by adding a randomly initialized linear projection on top of the context network into classes representing the vocabulary of the task (Baevski et al., 2019). For Librispeech, we have 29 tokens for character targets plus a word boundary token. Models are optimized by minimizing a CTC loss (Graves et al., 2006) and we apply a modified version of SpecAugment (Park et al., 2019) by masking to time-steps and channels during training which delays overfitting and significantly improves the final error rates, especially on the Libri-light subsets with few labeled examples.
4 Experimental Setup
As unlabeled data we consider the Librispeech corpus (Panayotov et al., 2015) without transcriptions containing 960 hours of audio (LS-960) or the audio data from LibriVox (LV-60k). For the latter we follow the pre-processing of (Kahn and others, 2020) resulting in 53.2k hours of audio. We fine-tune on five labeled data settings: 960 hours of transcribed Librispeech, the train-clean-100 subset comprising 100 hours (100 hours labeled), as well as the Libri-light limited resource training subsets originally extracted from Librispeech, these are train-10h (10 hours labeled), train-1h (1 hour labeled), train-10min (10 min labeled). We follow the evaluation protocol of Libri-light for these splits and evaluate on the standard Librispech dev-other/clean and test-clean/other sets.
We fine-tune the pre-trained models for phoneme recognition on the TIMIT dataset (Garofolo et al., 1993). It contains five hours of audio recordings with detailed phoneme labels. We use the standard train, dev and test split and follow the standard protocol of collapsing phone labels to 39 classes.
Models are implemented in fairseq Ott et al. (2019)
. The feature encoder contains seven blocks and the temporal convolutions in each block have 512 channels with strides (5,2,2,2,2,2,2) and kernel widths (10,3,3,3,3,2,2). This results in an encoder output frequency of 49 hz with a stride of about 20ms between each sample, and a receptive field of 400 input samples or 25ms of audio.
We experiment with two model configurations which use the same encoder architecture but differ in the Transformer setup: Base contains 12 transformer blocks, model dimension 768, inner dimension (FFN) 3,072 and 8 attention heads. Batches are built by cropping 250k audio samples, or 15.6sec, from each example. Crops are batched together to not exceed 1.4m samples per GPU and we train on a total of 64 V100 GPUs for 1.6 days Ott et al. (2018); the total batch size is 1.6h.
The Large model contains 24 transformer blocks with model dimension 1,024, inner dimension 4,096 and 16 attention heads. We crop 320K audio samples, or 20sec, with a limit of 1.2M samples per GPU and train on 128 V100 GPUs over 2.3 days for Librispeech and 5.2 days for LibriVox; the total batch size is 2.7h. We use dropout 0.1 in the Transformer, at the output of the feature encoder and the input to the quantization module. Layers are dropped at a rate of 0.05 for Base and 0.2 for Large Huang et al. (2016); Fan et al. (2020); there is no layer drop for LV-60k.
We optimize with Adam (Kingma and Ba, 2015), warming up the learning rate for the first 8% of updates to a peak of for Base and for Large, and then linearly decay it. Large trains for 250k updates, Base for 400k updates, and Large on LV-60k for 600k updates. We use weight for the diversity loss and for the feature penalty in Equation 2. For the quantization module we use and for both models, resulting in a theoretical maximum of 102.4k codewords. Entries are of size for Base amd for Large. The Gumbel softmax temperature is annealed from 2 to a minimum of 0.5 for Base and 0.1 for Large by a factor of 0.999995 at every update. The temperature in the contrastive loss (Equation 3) is set to . We set the feature encoder gradient scaling factor to for Librispech and for LibriVox. In the contrastive loss we use distractors. We choose the training checkpoint with the lowest on the validation set.
After pre-training we fine-tune the learned representations on labeled data and add a randomly initialized output layer on top of the Transformer to predict (Librispeech/Libri-light) or phonemes (TIMIT). For Libri-light, we train three seeds with two different learning rates (2e-5 and 3e-5) for all subsets and choose the configuration with lowest WER on dev-other subset decoded with the official 4-gram language model (LM) with beam 50 and fixed model weights (LM weight 2, word insertion penalty -1). For Base on the labeled 960h subset we use a learning rate of 1e-4.
We optimize with Adam and a tri-state rate schedule where the learning rate is warmed up for the first 10% of updates, held constant for the next 40% and then linearly decayed for the remainder. Base uses a batch size of 3.2m samples per GPU and we fine-tune on 8 GPUs, giving a total batch size of 1,600sec. Large
batches 1.28M samples on each GPU and we fine-tune on 24 GPUs, resulting in an effective batch size of 1920sec. For the first 10k updates only the output classifier is trained, after which the Transformer is also updated. The feature encoder is not trained during fine-tuning. We mask the feature encoder representations with a strategy similar to SpecAugment(Park et al., 2019) detailed in Appendix B.
4.4 Language Models and Decoding
We consider two types of language models (LM): a 4-gram model and a Transformer (Baevski and Auli, 2018) trained on the Librispeech LM corpus. The Transformer LM is identical to (Synnaeve et al., 2020) and contains 20 blocks, model dimension 1280, inner dimension 6144 and 16 attention heads. We tune the weights of the language model (interval ) and a word insertion penalty () via Bayesian optimization333https://github.com/facebook/Ax
: we run 128 trials with beam 500 for the 4-gram LM and beam 50 for the Transformer LM and choose the best set of weights according to performance on dev-other. Test performance is measured with beam 1,500 for the n-gram LM and beam 500 for the Transformer LM. We use the beam search decoder ofPratap et al. (2019).
|10 min labeled|
|Discrete BERT (Baevski et al., 2019)||LS-960||4-gram||15.7||24.1||16.3||25.2|
|Discrete BERT (Baevski et al., 2019)||LS-960||4-gram||8.5||16.4||9.0||17.6|
|Discrete BERT (Baevski et al., 2019)||LS-960||4-gram||5.3||13.2||5.9||14.1|
|Hybrid DNN/HMM (Lüscher et al., 2019)||-||4-gram||5.0||19.5||5.8||18.6|
|TTS data augm. (Laptev et al., 2020)||-||LSTM||4.3||13.5|
|Discrete BERT (Baevski et al., 2019)||LS-960||4-gram||4.0||10.9||4.5||12.1|
|Iter. pseudo-labeling Xu et al. (2020)||LS-860||4-gram+Transf.||5.0||8.72||5.37||9.51|
|Iter. pseudo-labeling Xu et al. (2020)||+LV-60k||4-gram+Transf.||3.19||6.14||3.72||7.11|
|Noisy student (Park et al., 2020)||LS-860||LSTM||3.9||8.8||4.2||8.6|
5.1 Low-Resource Labeled Data Evaluation
We first evaluate our pre-trained models in settings where the amount of labeled data is limited to get a sense of how the representations learned on unlabeled data can improve low resource settings. If a pre-trained model captures the structure of speech, then it should require few labeled examples to fine-tune it for speech recognition. The models are pre-trained on the audio data of either Librispeech (LS-960) or LibriVox (LV-60k) and most results are obtained by decoding with a Transformer language model (Transf.); Appendix C shows results with other language models.
The Large model pre-trained on LV-60k and fine-tuned on only 10 minutes of labeled data achieves a word error rate of 5.7/10.1 on the Librispeech clean/other test sets. Ten minutes of labeled data corresponds to just 48 recordings with an average length of 12.5 seconds. This demonstrates that ultra-low resource speech recognition is possible with self-supervised learning on unlabeled data. Our approach improves over previous pre-training work which did not learn quantized audio units jointly (Baevski et al., 2019), reducing WER by a about a third.
A recent iterative self-training approach (Park et al., 2020) represents the state of the art on the clean 100 hour subset of Librispeech but it requires multiple iterations of labeling, filtering, and re-training. Our approach is simpler: we pre-train on the unlabeled data and fine-tune on the labeled data. On the 100 hour subset of Librispeech, their method achieves WER 4.2/8.6 on test-clean/other which compares to WER 2.3/5.0 with the Large model in a like for like setup, a relative WER reduction of 45%/42%.
When the Large model uses an order of magnitude less labeled data (10h labeled), then it still achieves WER 3.2/6.1, an error reduction of 24%/29% relative to iterative self-training. Using only a single hour of labeled data, the same model achieves WER 3.9/7.6 which improves on both test-clean and test-other by 7%/12% - with two orders of magnitude less labeled data. We note that the Libri-light data splits contain both clean and noisy data leading to better accuracy on test-other compared to test-clean. Increasing model size reduces WER on all setups with the largest improvements on test-other (Base vs. Large both on LS-960) and increasing the amount of unlabeled training data also leads to large improvements (Large LS-960 vs. LV-60k).
5.2 High-Resource Labeled Data Evaluation on Librispeech
In this section we evaluate the performance when large quantities of labeled speech are available to assess the effectiveness of our approach in a high resource setup. Specifically, we fine-tune the same models as before on the full 960 hours of labeled Librispeech: Base and Large pre-trained on LS-960 as well as Large pre-trained on LV-60k.
Table 2 shows that our approach achieves WER 1.9/3.5 on test-clean/other. To our knowledge, this is the first time self-supervised learning achieves results competitive to the state of the art iterative semi-supervised methods in a high-resource labeled data setup. This is despite a weaker baseline architecture: supervised training of our architecture achieves WER 2.1/4.6 (Large - from scratch) compared to WER 1.9/4.1 for ContextNet (Han et al., 2020), the baseline architecture of the state of the art (Park et al., 2020).
We expect a better architecture to lead to improvements: we use a simple Tranfsormer with CTC which does not perform as well as seq2seq models (Synnaeve et al., 2020). The vocabulary of our acoustic model (characters) does not match the vocabulary of the LM (words) which delays feedback from the LM and is likely to be detrimental. Most recent work (Synnaeve et al., 2020; Xu et al., 2020; Han et al., 2020; Park et al., 2020) uses the better performing word pieces (Schuster and Nakajima, 2012) for both models. Moreover, our result is achieved without any data balancing such as Park et al. (2020). Finally, self-training is likely complimentary to pre-training and their combination may yield even better results. Appendix E presents a detailed error analysis of our pre-trained models in various labeled data setups.
|CTC Transf (Synnaeve et al., 2020)||-||CLM+Transf.||2.20||4.94||2.47||5.45|
|S2S Transf. (Synnaeve et al., 2020)||-||CLM+Transf.||2.10||4.79||2.33||5.17|
|Transf. Transducer (Zhang et al., 2020)||-||Transf.||-||-||2.0||4.6|
|ContextNet (Han et al., 2020)||-||LSTM||1.9||3.9||1.9||4.1|
|CTC Transf. + PL (Synnaeve et al., 2020)||LV-60k||CLM+Transf.||2.10||4.79||2.33||4.54|
|S2S Transf. + PL (Synnaeve et al., 2020)||LV-60k||CLM+Transf.||2.00||3.65||2.09||4.11|
|Iter. pseudo-labeling Xu et al. (2020)||LV-60k||4-gram+Transf.||1.85||3.26||2.10||4.01|
|Noisy student (Park et al., 2020)||LV-60k||LSTM||1.6||3.4||1.7||3.4|
|Large - from scratch||-||Transf.||1.7||4.3||2.1||4.6|
5.3 Phoneme Recognition on TIMIT
Next, we evaluate accuracy on TIMIT phoneme recognition by fine-tuning the pre-trained models on the labeled TIMIT training data. We fine-tune as for the 10 hour subset of Libri-light but do not use a language model. Table 3 shows that our approach can achieve a new state of the art on this dataset, reducing PER by a relative 23%/29% over the next best result on the dev/test sets. Appendix D shows an analysis of how the discrete latent speech representations related to phonemes.
|dev PER||test PER|
|CNN + TD-filterbanks (Zeghidour et al., 2018)||15.6||18.0|
|Li-GRU + fMLLR (Ravanelli et al., 2018)||–||14.9|
|wav2vec (Schneider et al., 2019)||12.9||14.7|
|vq-wav2vec (Baevski et al., 2020)||9.6||11.6|
|This work (no LM)|
A difference to previous work Baevski et al. (2020, 2019) is that we quantize the latent audio representations only for the contrastive loss, i.e., when latents are used as targets, but not when the latents are input to the Transformer network. We motivate this choice by an ablating for which we adopt a reduced training setup to increase experimental turn around: we pre-train Base on LS-960 for 250k updates with masking probability
, fine-tune on train-10h for 60k updates on a single GPU with 640k samples per batch, or 40 sec of speech audio. We report the average WER and standard deviation on the concatenation of dev-clean and dev-other (dev PER) for three seeds of fine-tuning.
Table 4 shows that our strategy of continuous inputs with quantized targets (Baseline) performs best. Continuous latent speech representations retain more information to enable better context representations and quantizing the target representations leads to more robust training. Quantizing the latents both in the input and the targets performs least well, and explains the lower performance of prior work (Baevski et al., 2020, 2019). Continuous targets reduce the effectiveness of self-supervised training since targets can capture detailed artifacts of the current sequence, e.g. speaker and background information, which make the task easier and prevent the model from learning general representations beneficial to speech recognition. The training accuracy of identifying the correct latent audio representation increases from 62% to 78.0% when switching from quantized to continuous targets. Continuous inputs and continuous targets perform second best but various attempts to improve it did not lead to better results (see Appendix F for this experiment and other ablations on various hyperparameters).
|Continuous inputs, quantized targets (Baseline)||7.97||0.02|
|Quantized inputs, quantized targets||12.18||0.41|
|Quantized inputs, continuous targets||11.18||0.16|
|Continuous inputs, continuous targets||8.58||0.08|
We presented wav2vec 2.0, a framework for self-supervised learning of speech representations which masks latent representations of the raw waveform and solves a contrastive task over quantized speech representations. Our experiments show the large potential of pre-training on unlabeled data for speech processing: when using only 10 minutes of labeled training data, or 48 recordings of 12.5 seconds on average, we achieve a WER of 5.7/10.1 on test-clean/other of Librispeech.
Our model achieves a new state of the art on the clean 100 hour Librispeech setup and outperforms the previous best result even when using 100 times less labeled data. The approach is also effective when large amounts of labeled data are available. We expect performance gains by switching to a seq2seq architecture and a word piece vocabulary.
There are around 7,000 languages in the world and many more dialects. However, for most of them no speech recognition technology exists since current systems require hundreds or thousands of hours of labeled data which is hard to collect for most languages. We have shown that speech recognition models can be built with very small amounts of annotated data at very good accuracy. We hope our work will make speech recognition technology more broadly available to many more languages and dialects.
We thank Tatiana Likhomanenko and Qiantong Xu for helpful discussion and their help with wav2letter integration.
- Layer normalization. arXiv. Cited by: §2.
- Learning representations by maximizing mutual information across views. In Proc. of NeurIPS, Cited by: §1.
- Effectiveness of self-supervised pre-training for speech recognition. arXiv abs/1911.03912. Cited by: §1, §1, §2, §2, §3.3, §5.1, §5.4, §5.4, Table 1.
- Adaptive input representations for neural language modeling. In Proc. of ICLR, Cited by: §4.4.
- Vq-wav2vec: self-supervised learning of discrete speech representations. In Proc. of ICLR, Cited by: §1, §1, §2, §2, §5.4, §5.4, Table 3.
- A simple framework for contrastive learning of visual representations. arXiv abs/2002.05709. Cited by: §1, §3.2.
Unsupervised speech representation learning using wavenet autoencoders. arXiv abs/1901.08810. Cited by: §1.
An unsupervised autoregressive model for speech representation learning. arXiv abs/1904.03240. Cited by: §1.
- BERT: pre-training of deep bidirectional transformers for language understanding. arXiv abs/1810.04805. Cited by: §1, §1, §2, §2, §3.
- The challenge of realistic music generation: modelling raw audio at scale. arXiv. Cited by: §3.2.
- Unsupervised acoustic unit discovery for speech synthesis using discrete latent-variable neural networks. arXiv abs/1904.07556. Cited by: §1.
- Reducing transformer depth on demand with structured dropout. In Proc. of ICLR, Cited by: Appendix B, §4.2.
- The DARPA TIMIT Acoustic-Phonetic Continuous Speech Corpus CDROM. Linguistic Data Consortium. Cited by: §4.1.
Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proc. of ICML, Cited by: §1, §3.3.
- Statistical theory of extreme values and some practical applications: a series of lectures. Vol. 33, US Government Printing Office. Cited by: §2.
ContextNet: improving convolutional neural networks for automatic speech recognition with global context. arXiv. Cited by: §5.2, §5.2, Table 2.
- Learning hierarchical discrete linguistic units from visually-grounded speech. In Proc. of ICLR, Cited by: §1.
- Momentum contrast for unsupervised visual representation learning. arXiv abs/1911.05722. Cited by: §1, §3.2.
- Data-efficient image recognition with contrastive predictive coding. arXiv abs/1905.09272. Cited by: §1.
- Gaussian error linear units (gelus). arXiv. Cited by: §2.
- Deep networks with stochastic depth. arXiv. Cited by: Appendix B, §4.2.
- Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proc. of AISTATS, Cited by: §3.2.
- Categorical reparameterization with gumbel-softmax. arXiv abs/1611.01144. Cited by: §1, §2.
- Product quantization for nearest neighbor search. IEEE Trans. Pattern Anal. Mach. Intell. 33 (1), pp. 117–128. Cited by: §2.
- Improving transformer-based speech recognition using unsupervised pre-training. arXiv abs/1910.09932. Cited by: §1, §1, §2.
- Libri-light: a benchmark for asr with limited or no supervision. In Proc. of ICASSP, Cited by: §4.1.
- Learning robust and multilingual speech representations. arXiv. Cited by: §1.
- Adam: A Method for Stochastic Optimization. In Proc. of ICLR, Cited by: §4.2.
- You do not need more data: improving end-to-end speech recognition by text-to-speech data augmentation. arXiv abs/2005.07157. Cited by: Table 1.
- Ethnologue: languages of the world, nineteenth edition. SIL International. Note: Online version: http://www.ethnologue.com Cited by: §1.
- Towards unsupervised speech recognition and synthesis with quantized speech representation learning. arXiv. Cited by: §1.
- RoBERTa: a robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Cited by: §2.
- RWTH asr systems for librispeech: hybrid vs attention. In Interspeech 2019, Cited by: Table 1.
- A* sampling. In Advances in Neural Information Processing Systems, pp. 3086–3094. Cited by: §2.
- Self-supervised learning of pretext-invariant representations. arXiv. Cited by: §1.
- Transformers with convolutional context for ASR. arXiv abs/1904.11660. Cited by: §2.
- Fairseq: a fast, extensible toolkit for sequence modeling. In Proc. of NAACL System Demonstrations, Cited by: §4.2.
Scaling neural machine translation. In Proc. of WMT, Cited by: §4.2.
- Librispeech: an asr corpus based on public domain audio books. In Proc. of ICASSP, pp. 5206–5210. Cited by: §4.1.
- SpecAugment: a simple data augmentation method for automatic speech recognition. In Proc. of Interspeech, Cited by: Appendix B, §3.3, §4.3.
- Improved noisy student training for automatic speech recognition. arXiv abs/2005.09629. Cited by: §1, §5.1, §5.2, §5.2, Table 1, Table 2.
- Deep contextualized word representations. In Proc. of ACL, Cited by: §1.
Wav2Letter++: a fast open-source speech recognition system. In Proc. of ICASSP, Cited by: §4.4.
- Improving language understanding by generative pre-training. Note: https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf Cited by: §1.
Light gated recurrent units for speech recognition. IEEE Transactions on Emerging Topics in Computational Intelligence 2 (2), pp. 92–102. Cited by: Table 3.
- Unsupervised pretraining transfers well across languages. arXiv abs/2002.02848. Cited by: §1.
- Wav2vec: unsupervised pre-training for speech recognition. In Proc. of Interspeech, Cited by: §1, Table 3.
- Japanese and korean voice search. In Proc. of ICASSP, Cited by: §5.2.
End-to-end ASR: from Supervised to Semi-Supervised Learning with Modern Architectures. arXiv abs/1911.08460. Cited by: §4.4, §5.2, Table 2.
- VQVAE unsupervised unit discovery and multi-scale code2spec inverter for zerospeech challenge 2019. arXiv 1905.11449. Cited by: §1.
- Representation learning with contrastive predictive coding. arXiv abs/1807.03748. Cited by: §3.2.
- Neural discrete representation learning. In Advances in Neural Information Processing Systems, pp. 6306–6315. Cited by: §1, §1.
- Attention is all you need. In Proc. of NIPS, Cited by: §2.
- Unsupervised pre-training of bidirectional speech encoders via masked reconstruction. arXiv. Cited by: §1.
- Pay less attention with lightweight and dynamic convolutions. In Proc. of ICLR, Cited by: §2.
- Iterative pseudo-labeling for speech recognition. arXiv. Cited by: §5.2, Table 1, Table 2.
- Learning filterbanks from raw speech for phone recognition. In Proc. of ICASSP, Cited by: Table 3.
- Transformer transducer: a streamable speech recognition model with transformer encoders and rnn-t loss. arXiv. Cited by: Table 2.
Appendix A Masking distribution
When choosing which time-steps to mask, each latent speech representation in an utterance is considered a candidate starting time-step with probability where is the length of each masked span starting from the respective time step; both are hyper-parameters. Sampled starting time steps are expanded to length and spans can overlap.
For a 15 sec long audio sample, the average mask length is 14.7 time-steps, corresponding to 299ms of audio, with a median of 10 time-steps, and a maximum of about 100 time steps; about 49% of all time-steps in the sample will be masked. A plot of the corresponding mask length distribution is shown in Figure 2 and an ablation of and as well as the effect of other masking strategies is shown in Table 5. Reducing results in increased prediction accuracy for the self-supervised but the task becomes trivial when spans with length one are masked, leading to poor performance on downstream speech recognition tasks. We also consider other masking strategies: w/o overlap uniform(,) samples for each starting index a span length from interval to and masks the subsequent time-steps taking care not to overlap with existing spans; poisson() and normal(, ) sample
from Poisson and normal distributions.
|Mask w/o overlap, uniform(1,31)||8.39||0.02|
|Mask w/o overlap, uniform(10,30)||9.17||0.05|
|Mask w/o overlap, poisson(15)||8.13||0.04|
|Mask w/o overlap, normal(15, 10)||8.37||0.03|
|Mask w/o overlap, length 10||9.15||0.02|
|Mask w/o overlap, length 15||9.43||0.26|
Appendix B Fine-tuning Setup
During fine-tuning we apply a masking strategy to the feature encoder outputs similar to SpecAugment [Park et al., 2019]: we randomly choose a number of starting time steps for which a span of ten subsequent time-steps is replaced with a mask embedding; spans may overlap and we use the same masked time step embedding as during pre-training. We also mask channels by choosing a number of channels as starting indices and then expand each one to cover the subsequent 64 channels. Spans may overlap and the selected channel spans are set to zero value. We use LayerDrop [Huang et al., 2016, Fan et al., 2020] at a rate of 0.05 for Base and 0.1 for Large during fine-tuning.
Table 6 summarizes the fine-tuning hyper-parameter settings used for the different labeled data setup. Table 7 contains the decoding parameters used for final evaluations of the various labeled data setups.
|timestep mask prob.||channel mask prob.||updates|
|4gram LM weight||4gram word insert.||TransLM weight||TransLM word insert.|
Appendix C Full results for Libri-light and Librispeech
|10 min labeled|
|Large - from scratch||-||None||2.8||7.6||2.6||5.8|
Appendix D Analysis of Discrete Latent Speech Representations
Next, we investigate whether the discrete latent speech representations learned by the quantizer relate to phonetic information: Using Large pre-trained on LV-60k and without any fine-tuning, we compute the discrete latents for the training data of TIMIT and compute the co-occurrence between human annotated phonemes and the latents. Ties are broken by choosing the phoneme which is most represented in the receptive field of . The training data contains 3696 utterances of average length 13.6 sec, or 563k discrete latents.
Figure 3 plots and shows that many discrete latents appear to specialize in specific phonetic sounds. The silence phoneme (bcl) represents 22% of all human annotated speech data and is therefore also modeled by many different latents.
Appendix E Speech Recognition Error Analysis
In this section we study the most common errors our models make when fine-tuned on different amounts of labeled data (Table 10). We also show transcriptions of a few relatively challenging utterances from the dev-clean subset of Librispeech (Table 11).
We consider models with no lexicon or no language model decoding, marked None inTable 8: Larger capacity decreases error rates: Large on LS-960 improves the word error rate on dev-clean from 46.1 to 43 compared to Base. Increasing the amount of unlabeled training data further decreases the error rate to 33.8 for Large on LS-960.
In the ten minute labeled data setup, the model is still able to recognize basic units of speech: Table 10 shows that most errors are around spelling of words, e.g., omitting silent characters such as could coud, know now, or ignoring repeated letters such as still stil, little litle. The Large LV-60k model achieves WER 33.8 on dev-clean, and when a lexicon constrains the search, spellings outside the vocabulary are pruned and the next closest word in the vocabulary is chosen instead. This lowers WER to 27.8. Adding a Transformer language model enables to choose more likely pronunciations during the search and gives a large WER improvement to 6.0.
The ten minute models without lexicon and language model tend to spell words phonetically and omit repeated letters, e.g., will wil (Table 10). Spelling errors decrease with more labeled data: with one hour of labeled data, slightly less common words move into the list of the most frequent errors, e.g., heaven and food are spelled phonetically. At ten hours, top errors include articles, e.g., a, the which are a common source of errors in speech recognition in general. There are also alternative spellings, color vs. colour as well as relatively rare words including person names, still spelled phonetically, e.g., phoebe feeby.
At 100 hours, person names dominate the most frequent errors: phoebe phebe, along with incorrect spacing anyone any one, awhile a while. Finally at 960 hours the word error rate falls to 2% and top errors are mostly articles, incorrect splits, and some very rare words or names such as deucalion or gryce.
The “from scratch” 960 hour model has a similar word error rate as the 100 hour pre-trained model and displays a similar pattern of errors.
The pre-trained speech representations can be easily adapted to recognize specific sounds while fine-tuning grounds these representations to the actual spelling.
|10m Large LV-60k||1h Large LV-60k||10h Large LV-60k|
|all al (181)||too to (26)||in and (15)|
|are ar (115)||until untill (24)||a the (11)|
|will wil (100)||new knew (22)||o oh (10)|
|you yo (90)||door dor (18)||and in (9)|
|one on (89)||says sais (18)||mode mod (9)|
|two to (81)||soul sol (17)||ursus ersus (9)|
|well wel (80)||bread bred (16)||tom tome (8)|
|been ben (73)||poor pore (16)||randal randol (7)|
|upon apon (73)||a the (13)||the a (7)|
|good god (67)||either ither (13)||color colour (6)|
|see se (66)||food fud (13)||flour flower (6)|
|we whe (60)||doubt dout (12)||phoebe feeby (6)|
|little litle (54)||earth erth (12)||an and (5)|
|great grate (53)||led lead (12)||cucumbers cucombers (5)|
|your yor (53)||sea see (12)||egg eg (5)|
|could coud (51)||thee the (12)||macklewain macklewaine (5)|
|here hear (51)||tom tome (12)||magpie magpi (5)|
|know now (45)||add ad (11)||milner millner (5)|
|there ther (45)||good god (11)||stacy staci (5)|
|three thre (45)||heaven heven (11)||trevelyan trevellion (5)|
|still stil (42)||mary marry (11)||verloc verlock (5)|
|off of (40)||randal randel (11)||ann an (4)|
|don’t dont (37)||answered ansered (10)||anyone one (4)|
|shall shal (36)||blood blod (10)||apartment appartment (4)|
|little litl (35)||bozzle bosel (10)||basin bason (4)|
|100h Large LV-60k||960h Large LV-60k||960h Large from scratch|
|a the (13)||a the (12)||and in (20)|
|and in (10)||and in (9)||a the (16)|
|in and (10)||macklewain mackelwaine (7)||in and (13)|
|o oh (8)||in and (6)||the a (10)|
|minnetaki minnitaki (7)||o oh (6)||in an (8)|
|randal randall (7)||bozzle bosell (5)||and an (5)|
|christie cristy (6)||criss chris (5)||clarke clark (4)|
|macklewain mackelwane (6)||bozzle bosel (4)||grethel gretel (4)|
|randal randoll (6)||clarke clark (4)||macklewain mackelwaine (4)|
|bozzle bosall (5)||colored coloured (4)||this the (4)|
|kaliko calico (5)||grethel gretel (4)||an and (3)|
|trevelyan trevelian (5)||lige lyge (4)||anyone one (3)|
|an and (4)||the a (4)||bozzle basell (3)|
|and an (4)||and an (3)||buns bunds (3)|
|anyone one (4)||ann marianne (3)||carrie carry (3)|
|bozzle bozall (4)||butte bute (3)||criss chris (3)|
|clarke clark (4)||color colour (3)||he’s is (3)|
|gryce grice (4)||deucalion ducalion (3)||his is (3)|
|i’m am (4)||forcemeat meat (3)||honor honour (3)|
|in ind (4)||gryce grice (3)||lattimer latimer (3)|
|letty lettie (4)||honor honour (3)||millet mellet (3)|
|phoebe phebe (4)||kearny kirney (3)||pyncheon pension (3)|
|the a (4)||nuova noiva (3)||tad ted (3)|
|ann anne (3)||thing anything (3)||thing anything (3)|
|awhile while (3)||this the (3)||trevelyan trevelian (3)|
|Reference||i’m mister christopher from london|
|10m LV-60k||IM mister CRESTIFER FROME LUNDEN|
|1h LV-60k||IM mister CRISTIFFHER from LOUNDEN|
|10h LV-60k||i’m mister CHRYSTEPHER from london|
|100h LV-60k||i’m mister christopher from london|
|960h LV-60k||i’m mister christopher from london|
|960h scratch||I MISSTER christopher from london|
|Reference||il popolo e una bestia|
|10m LV-60k||ILPOPULAR ONABESTIA|
|1h LV-60k||O POPOLAONABASTIA|
|10h LV-60k||U POPULAONABASTIAR|
|100h LV-60k||O POPALOON A BASTYA|
|960h LV-60k||YOU’LL POP A LAWYE ON A BAISTYE|
|960h scratch||OL POPALOY ON ABESTIA|
|Reference||he smelt the nutty aroma of the spirit|
|10m LV-60k||he SMELTD the NUDY aroma of the spirit|
|1h LV-60k||he SMELTD the NUDDY ARROMA of the spirit|
|10h LV-60k||he smelt the NUDDY ERROMA of the spirit|
|100h LV-60k||he smelt the NUDDY aroma of the spirit|
|960h LV-60k||he smelt the NUTTIE aroma of the spirit|
|960h scratch||he smelt the nutty EROMA of the spirit|
|Reference||phoebe merely glanced at it and gave it back|
|10m LV-60k||FEABY MEARLY glanced at it and gave it BAK|
|1h LV-60k||FIEABY merely glanced at it and gave it back|
|10h LV-60k||FEEBY merely glanced at it and gave it back|
|100h LV-60k||BEBE merely glanced at it and gave it back|
|960h LV-60k||phoebe merely glanced at it and gave it back|
|960h scratch||phoebe merely glanced at it and gave it back|
|Reference||sauterne is a white bordeaux a strong luscious wine the best known varieties being|
|10m LV-60k||SULTERIN is a white BORDOE a strong LUCHOUS WIN the best NOWN VERIATYS being|
|1h LV-60k||CLTEREN is a white BORDO a strong LUCHIOUS wine the best known VERIETIES being|
|10h LV-60k||SOTERN is a white BOURDO a strong LUCIOUS wine the best known VORIETIES being|
|100h LV-60k||SOTERN is a white BORDAUX a strong LUCIOUS wine the best known varieties being|
|960h LV-60k||SOTERN is a white bordeaux a strong luscious wine the best known varieties being|
|960h scratch||SOTERAN is a white bordeaux a strong luscious wine the best known varieties being|
|Reference||i happen to have mac connell’s box for tonight or there’d be no chance of our getting places|
|10m LV-60k||i HAPEND to have MECONALES BOXS for TONIT ORE THIRLD be no chance of OR GETING places|
|1h LV-60k||i happen to have MACCONNEL’S BOCXS for tonight or TE’ELD be no chance of our getting places|
|10h LV-60k||i HAPPENED to have MUKONNEL’S box for tonight or THERED be no chance of our getting places|
|100h LV-60k||i HAPPENED to have MC CONNEL’S box for TO NIGHT or there’d be no chance of our getting places|
|960h LV-60k||i happen to have MC CONALL’S box for TO NIGHT or there’d be no chance of our getting places|
|960h scratch||i HAPPENE to have MACONEL’S box for TO NIGHT or there’d be no chance of our getting places|
Appendix F Ablations
Table 12 ablates various hyperparameter choices of our architecture. The setup for the baseline model is described in § 5.4. First, we tried to improve the continuous input and continuous target model (§ 5.4) by adding an MLP on top of the continuous target representation and we also tried to use a separate set of encoder parameters for the representations used as input and targets (Separate encoders). Both did not lead to meaningful improvements.
Increasing the receptive field size from 25ms to 30ms had little effect. Setting the diversity penalty weight () too low results in lower codebook usage and lower performance. Setting it too high leads to slight instability. Removing the L2 penalty over the feature encoder outputs decreases performance (). Doubling the number of relative positional embeddings to 256 also did not help. Stopping gradients from the quantizer to the encoder shows that the encoder requires training signal from the quantizer as well.
|Baseline (, , )||7.97||0.02|
|Continuous inputs, continuous targets||8.58||0.08|
|+ MLP on targets||8.51||0.05|
|+ Separate encoders||8.90||0.01|
|receptive field 30ms||7.99||0.06|
|Feature L2 penalty||8.38||0.07|
|Conv pos emb, kernel 256||8.14||0.05|
|No gradient to encoder from quantizer||8.41||0.08|
|same utterance + from batch||8.79||0.06|
|Sample negatives from any time step||8.07||0.02|
|No Gumbel noise||8.73||0.42|
|Predict exactly time steps from edges|
Next, increasing the number of negatives did not result in better performance () and sampling negatives from the entire batch of utterances hurt performance, likely because candidates from other utterances are easy to distinguish. Sampling negatives from any time step in the utterance, masked or unmasked, does not help and is more computationally expensive. Gumbel noise is important and increasing the number of codebooks did not result in better performance.
We also investigated predicting only time steps immediately next to the last unmasked time step for each span. This enables to better control the difficulty of the pre-training task. Given the leftmost or rightmost unmasked time step next to a masked span, we compute the contrastive loss only for the first masked time steps next to these unsmasked spans. Predicting only up to one time step performs poorly because there is little training signal from each utterance and predicting more time steps performs better but does not significantly outperform predicting all masked time steps. Increasing the number of training updates helps but this increases training time.