We sometimes observe monotonically decreasing cross-attention weights in our Conformer-based global attention-based encoder-decoder (AED) models, negatively affecting performance compared to monotonically increasing attention weights. Further investigation shows that the Conformer encoder reverses the sequence in the time dimension. We analyze the initial behavior of the decoder cross-attention mechanism and find that it encourages the Conformer encoder self-attention to build a connection between the initial frames and all other informative frames. Furthermore, we show that, at some point in training, the self-attention module of the Conformer starts dominating the output over the preceding feed-forward module, which then only allows the reversed information to pass through. We propose methods and ideas of how this flipping can be avoided and investigate a novel method to obtain label-frame-position alignments by using the gradients of the label log probabilities w.r.t. the encoder input frames.

We study a streamable attention-based encoder-decoder model in which either the decoder, or both the encoder and decoder, operate on pre-defined, fixed-size windows called chunks. A special end-of-chunk (EOC) symbol advances from one chunk to the next chunk, effectively replacing the conventional end-of-sequence symbol. This modification, while minor, situates our model as equivalent to a transducer model that operates on chunks instead of frames, where EOC corresponds to the blank symbol. We further explore the remaining differences between a standard transducer and our model. Additionally, we examine relevant aspects such as long-form speech generalization, beam size, and length normalization. Through experiments on Librispeech and TED-LIUM-v2, and by concatenating consecutive sequences for long-form trials, we find that our streamable model maintains competitive performance compared to the non-streamable variant and generalizes very well to long-form speech.

We introduce a novel segmental-attention model for automatic speech recognition. We restrict the decoder attention to segments to avoid quadratic runtime of global attention, better generalize to long sequences, and eventually enable streaming. We directly compare global-attention and different segmental-attention modeling variants. We develop and compare two separate time-synchronous decoders, one specifically taking the segmental nature into account, yielding further improvements. Using time-synchronous decoding for segmental models is novel and a step towards streaming applications. Our experiments show the importance of a length model to predict the segment boundaries. The final best segmental-attention model using segmental decoding performs better than global-attention, in contrast to other monotonic attention approaches in the literature. Further, we observe that the segmental model generalizes much better to long sequences of up to several minutes.

Attention-based encoder-decoder (AED) models learn an implicit internal language model (ILM) from the training transcriptions. The integration with an external LM trained on much more unpaired text usually leads to better performance. A Bayesian interpretation as in the hybrid autoregressive transducer (HAT) suggests dividing by the prior of the discriminative acoustic model, which corresponds to this implicit LM, similarly as in the hybrid hidden Markov model approach. The implicit LM cannot be calculated efficiently in general and it is yet unclear what are the best methods to estimate it. In this work, we compare different approaches from the literature and propose several novel methods to estimate the ILM directly from the AED model. Our proposed methods outperform all previous approaches. We also investigate other methods to suppress the ILM mainly by decreasing the capacity of the AED model, limiting the label context, and also by training the AED model together with a pre-existing LM.

We present our transducer model on Librispeech. We study variants to include an external language model (LM) with shallow fusion and subtract an estimated internal LM. This is justified by a Bayesian interpretation where the transducer model prior is given by the estimated internal LM. The subtraction of the internal LM gives us over 14% relative improvement over normal shallow fusion. Our transducer has a separate probability distribution for the non-blank labels which allows for easier combination with the external LM, and easier estimation of the internal LM. We additionally take care of including the end-of-sentence (EOS) probability of the external LM in the last blank probability which further improves the performance. All our code and setups are published.

With the advent of direct models in automatic speech recognition (ASR), the formerly prevalent frame-wise acoustic modeling based on hidden Markov models (HMM) diversified into a number of modeling architectures like encoder-decoder attention models, transducer models and segmental models (direct HMM).While transducer models stay with a frame-level model definition, segmental models are defined on the level of label segments directly.While (soft-)attention-based models avoid explicit alignment, transducer and segmental approach internally do model alignment, either by segment hypotheses or, more implicitly, by emitting so-called blank symbols.In this work, we prove that the widely used class of RNN-Transducer models and segmental models (direct HMM) are equivalent and therefore show equal modeling power.It is shown that blank probabilities translate into segment length probabilities and vice versa.In addition, we provide initial experiments investigating decoding and beam-pruning, comparing time-synchronous and label-/segment-synchronous search strategies and their properties using the same underlying model.

In this paper, we study a simple yet elegant latent variable attention model for automatic speech recognition (ASR) which enables an integration of attention sequence modeling into the direct hidden Markov model (HMM) concept. We use a sequence of hidden variables that establishes a mapping from output labels to input frames. Inspired by the direct HMM model, we assume a decomposition of the label sequence posterior into emission and transition probabilities using zero-order assumption and incorporate both Transformer and LSTM attention models into it. The method keeps the explicit alignment as part of the stochastic model and combines the ease of the end-to-end training of the attention model as well as an efficient and simple beam search. To study the effect of the latent model, we qualitatively analyze the alignment behavior of the different approaches. Our experiments on three ASR tasks show promising results in WER with more focused alignments in comparison to the attention models.

Long short-term memory (LSTM) networks are the dominant architecture for large vocabulary continuous speech recognition (LVCSR) acoustic modeling due to their good performance. However, LSTMs are hard to tune and computationally expensive. To build a system with lower computational costs and which allows online streaming applications, we explore convolutional neural networks (CNN). To the best of our knowledge there is no overview on CNN hyper-parameter tuning for LVCSR in the literature, so we present our results explicitly. Apart from recognition performance, we focus on the training and evaluation speed and provide a time-efficient setup for CNNs. We faced an overfitting problem in training and solved it with data augmentation, namely SpecAugment. The system achieves results competitive with the top LSTM results. We significantly increased the speed of CNN in training and decoding approaching the speed of the offline LSTM.

Recent advances in text-to-speech (TTS) led to the development of flexible multi-speaker end-to-end TTS systems. We extend state-of-the-art attention-based automatic speech recognition (ASR) systems with synthetic audio generated by a TTS system trained only on the ASR corpora itself. ASR and TTS systems are built separately to show that text-only data can be used to enhance existing end-to-end ASR systems without the necessity of parameter or architecture changes. We compare our method with language model integration of the same text data and with simple data augmentation methods like SpecAugment and show that performance improvements are mostly independent. We achieve improvements of up to 33% relative in word-error-rate (WER) over a strong baseline with data-augmentation in a low-resource environment (LibriSpeech-100h), closing the gap to a comparable oracle experiment by more than 50%. We also show improvements of up to 5% relative WER over our most recent ASR baseline on LibriSpeech-960h.

Training deep neural networks is often challenging in terms of training stability. It often requires careful hyperparameter tuning or a pretraining scheme to converge. Layer normalization (LN) has shown to be a crucial ingredient in training deep encoder-decoder models. We explore various LN long short-term memory (LSTM) recurrent neural networks (RNN) variants by applying LN to different parts of the internal recurrency of LSTMs. There is no previous work that investigates this. We carry out experiments on the Switchboard 300h task for both hybrid and end-to-end ASR models and we show that LN improves the final word error rate (WER), the stability during training, allows to train even deeper models, requires less hyperparameter tuning, and works well even without pre-training. We find that applying LN to both forward and recurrent inputs globally, which we denoted by Global Joined Norm variant, gives a 10% relative improvement in WER.

The RNN transducer is a promising end-to-end model candidate. We compare the original training criterion with the full marginalization over all alignments, to the commonly used maximum approximation, which simplifies, improves and speeds up our training. We also generalize from the original neural network model and study more powerful models, made possible due to the maximum approximation. We further generalize the output label topology to cover RNN-T, RNA and CTC. We perform several studies among all these aspects, including a study on the effect of external alignments. We find that the transducer model generalizes much better on longer sequences than the attention model. Our final transducer model outperforms our attention model on Switchboard 300h by over 6% relative WER.

While recurrent neural networks can motivate cross-sentence language modeling and its application to automatic speech recognition (ASR), corresponding modifications of the training method for that end are rarely discussed. In fact, even more generally, the impact of training sequence construction strategy in language modeling for different evaluation conditions is typically ignored. In this work, we revisit this basic but fundamental question. We train language models based on long short-term memory recurrent neural networks and Transformers using various types of training sequences and study their robustness with respect to different evaluation modes. Our experiments on 300h Switchboard and Quaero English datasets show that models trained with back-propagation over sequences consisting of concatenation of multiple sentences with state carry-over across sequences effectively outperform those trained with the sentence-level training, both in terms of perplexity and word error rates for cross-utterance ASR.

We present competitive results using a Transformer encoder-decoder-attention model for end-to-end speech recognition needing less training time compared to a similarly performing LSTM model. We observe that the Transformer training is in general more stable compared to the LSTM, although it also seems to overfit more, and thus shows more problems with generalization. We also find that two initial LSTM layers in the Transformer encoder provide a much better positional encoding. Data-augmentation, a variant of SpecAugment, helps to improve both the Transformer by 33% and the LSTM by 15% relative. We analyze several pretraining and scheduling schemes, which is crucial for both the Transformer and the LSTM models. We improve our LSTM model by additional convolutional layers. We perform our experiments on Lib-riSpeech 1000h, Switchboard 300h and TED-LIUM-v2 200h, and we show state-of-the-art performance on TED-LIUM-v2 for attention based end-to-end models. We deliberately limit the training on LibriSpeech to 12.5 epochs of the training data for comparisons, to keep the results of practical interest, although we show that longer training time still improves more. We publish all the code and setups to run our experiments.

Attention-based sequence-to-sequence models have shown promising results in automatic speech recognition. Using these architectures, one-dimensional input and output sequences are related by an attention approach, thereby replacing more explicit alignment processes, like in classical HMM-based modeling. In contrast, here we apply a novel two-dimensional long short-term memory (2DLSTM) architecture to directly model the input/output relation between audio/feature vector sequences and word sequences. The proposed model is an alternative model such that instead of using any type of attention components, we apply a 2DLSTM layer to assimilate the context from both input observations and output transcriptions. The experimental evaluation on the Switchboard 300h automatic speech recognition task shows word error rates for the 2DLSTM model that are competitive to end-to-end attention-based model.

We explore deep autoregressive Transformer models in language modeling for speech recognition. We focus on two aspects. First, we revisit Transformer model configurations specifically for language modeling. We show that well configured Transformer models outperform our baseline models based on the shallow stack of LSTM recurrent neural network layers. We carry out experiments on the open-source LibriSpeech 960hr task, for both 200K vocabulary word-level and 10K byte-pair encoding subword-level language modeling. We apply our word-level models to conventional hybrid speech recognition by lattice rescoring, and the subword-level models to attention based encoder-decoder models by shallow fusion. Second, we show that deep Transformer language models do not require positional encoding. The positional encoding is an essential augmentation for the self-attention mechanism which is invariant to sequence ordering. However, in autoregressive setup, as is the case for language modeling, the amount of information increases along the position dimension, which is a positional signal by its own. The analysis of attention weights shows that deep autoregressive self-attention models can automatically make use of such positional information. We find that removing the positional encoding even slightly improves the performance of these models.

We present state-of-the-art automatic speech recognition (ASR) systems employing a standard hybrid DNN/HMM architecture compared to an attention-based encoder-decoder design for the LibriSpeech task. Detailed descriptions of the system development, including model design, pretraining schemes, training schedules, and optimization approaches are provided for both system architectures. Both hybrid DNN/HMM and attention-based systems employ bi-directional LSTMs for acoustic modeling/encoding. For language modeling, we employ both LSTM and Transformer based architectures. All our systems are built using RWTHs open-source toolkits RASR and RETURNN. To the best knowledge of the authors, the results obtained when training on the full LibriSpeech training set, are the best published currently, both for the hybrid DNN/HMM and the attention-based systems. Our single hybrid system even outperforms previous results obtained from combining eight single systems. Our comparison shows that on the LibriSpeech 960h task, the hybrid DNN/HMM system outperforms the attention-based system by 15% relative on the clean and 40% relative on the other test sets in terms of word error rate. Moreover, experiments on a reduced 100h-subset of the LibriSpeech training corpus even show a more pronounced margin between the hybrid DNN/HMM and attention-based architectures.

Sequence-to-sequence attention-based models on subword units allow simple open-vocabulary end-to-end speech recognition. In this work, we show that such models can achieve competitive results on the Switchboard 300h and LibriSpeech 1000h tasks. In particular, we report the state-of-the-art word error rates (WER) of 3.54% on the dev-clean and 3.82% on the test-clean evaluation subsets of LibriSpeech. We introduce a new pretraining scheme by starting with a high time reduction factor and lowering it during training, which is crucial both for convergence and final performance. In some experiments, we also use an auxiliary CTC loss function to help the convergence. In addition, we train long short-term memory (LSTM) language models on subword units. By shallow fusion, we report up to 27% relative improvements in WER over the attention baseline without a language model.

In this work we release our extensible and easily configurable neural network training software. It provides a rich set of functional layers with a particular focus on efficient training of recurrent neural network topologies on multiple GPUs. The source of the software package is public and freely available for academic research purposes and can be used as a framework or as a standalone tool which supports a flexible configuration. The software allows to train state-of-the-art deep bidirectional long short-term memory (LSTM) models on both one dimensional data like speech or two dimensional data like handwritten text and was used to develop successful submission systems in several evaluation campaigns.

Recent experiments show that deep bidirectional long short-term memory (BLSTM) recurrent neural network acoustic models outperform feedforward neural networks for automatic speech recognition (ASR). However, their training requires a lot of tuning and experience. In this work, we provide a comprehensive overview over various BLSTM training aspects and their interplay within ASR, which has been missing so far in the literature. We investigate on different variants of optimization methods, batching, truncated backpropagation, and regularization techniques such as dropout, and we study the effect of size and depth, training models of up to 10 layers. This includes a comparison of computation times vs. recognition performance. Furthermore, we introduce a pretraining scheme for LSTMs with layer-wise construction of the network showing good improvements especially for deep networks. The experimental analysis mainly was performed on the Quaero task, with additional results on Switchboard. The best BLSTM model gave a relative improvement in word error rate of over 15% compared to our best feed-forward baseline on our Quaero 50h task. All experiments were done using RETURNN and RASR, RWTH's extensible training framework for universal recurrent neural networks and ASR toolkit. The training configuration files are publicly available.

It has been shown that sequence-discriminative training can improve the performance for large vocabulary continuous speech recognition. Our main contribution is a novel method for reducing the computation time of any sort of sequence training while only slightly decreasing the overall performance. The method allows to parallelize the forward propagation through the network, the loss and loss gradient calculation which will provide a frame-wise error signal, and an independent forward and back propagation using that error signal. That last step can be calculated in a frame-wise manner and thus allows to use frame chunking to further improve the runtime. The loss calculation can itself be parallelized over many sequences. In addition to several experiments which outline the runtime gains, we also provide a convergence proof sketch. We extend on the research of sequence training of bidirectional long-short term memory ((B)LSTM) networks and provide an overview and comparison over different criteria. We have published all the code as part of our RETURNN and RASR framework including our training setup configurations.

We formulate a generalized hybrid HMM-NN training procedure using the full-sum over the hidden state-sequence and identify CTC as a special case of it.We present an analysis of the alignment behavior of such a training procedure and explain the strong localization of label output behavior of full-sum training (also referred to as peaky or spiky behavior).We show how to avoid that behavior by using a state prior.We discuss the temporal decoupling between output label position/time-frame, and the corresponding evidence in the input observations when this is trained with BLSTM models.We also show a way how to overcome this by jointly training a FFNN.We implemented the Baum-Welch alignment algorithm in CUDA to be able to do fast soft realignments on GPU.We have published this code along with some of our experiments as part of RETURNN, RWTH's extensible training framework for universal recurrent neural networks.We finish with experimental validation of our study on WSJ and Switchboard.

Recurrent neural networks that can be trained end-to-end on sequence learning tasks provide promising benefits over traditional recognition systems. In this paper, we demonstrate the application of an attention-based long short-term memory decoder network for offline handwriting recognition and analyze the segmentation, classification and decoding errors produced by the model. We further extend the decoding network by a bidirectional topology together with an integrated length estimation procedure and show that it is superior to unidirectional decoder networks. Results are presented for the word and text line recognition tasks of the RIMES handwriting recognition database. The software used in the experiments is freely available for academic research purposes.

Online-Recognition requires the acoustic model to provide posterior probabilities after a limited time delay given the online input audio data.This necessitates unidirectional modeling and the standard solution is to use unidirectional long short-term memory (LSTM) recurrent neural networks (RNN) or feedforward neural networks (FFNN).It is known that bidirectional LSTMs are more powerful and perform better than unidirectional LSTMs.To demonstrate the performance difference, we start by comparing several different bidirectional and unidirectional LSTM topologies.Furthermore, we apply a modification to bidirectional RNNs to enable online-recognition by moving a window over the input stream and perform one forwarding through the RNN on each window.Then, we combine the posteriors of each forwarding and we renormalize them.We show in experiments that the performance of this online-enabled bidirectional LSTM performs as good as the offline bidirectional LSTM and much better than the unidirectional LSTM.