Towards End-To-End Speech Recognition with Recurrent Neural Networks

article2014ICML2,307 citations

Presents an end-to-end speech recognition framework combining deep bidirectional LSTMs and Connectionist Temporal Classification to transcribe audio directly into text without phonetic alignments, achieving competitive word error rates by optimizing directly for sequence transcription loss.

Listen

Traditional automated speech recognition systems rely on complex, multi-stage pipelines that require phonetic dictionaries, alignment steps, and separate acoustic and language models. This complexity creates engineering bottlenecks and mismatches between the training objectives and final transcription accuracy.

The article evaluates whether an end-to-end speech recognition system can directly transcribe audio into text using a recurrent neural network without requiring intermediate phonetic representations or pre-aligned data.

The authors designed a deep bidirectional recurrent network utilizing connectionist temporal classification and introduced a novel objective function to directly optimize the word error rate. The system was evaluated on the standard Wall Street Journal audio dataset under varying training set sizes (14 hours and 81 hours) and across different decoding configurations ranging from no language model to trigram models, benchmarked against a standard baseline system.

The key findings demonstrate that on the full 81-hour dataset without any external language model or dictionary, the proposed system achieved a word error rate of 27.3%, significantly outperforming the baseline configuration. Adding a simple word lexicon reduced the error rate to 21.9%, and incorporating a trigram language model brought it down to 8.2%, nearly matching the baseline pipeline's 7.8%. Crucially, combining the neural network model with the traditional baseline achieved the best overall result, lowering the word error rate to 6.7%.

These results indicate that end-to-end neural architectures can substantially streamline the development of speech recognition systems, reducing the labor and domain expertise required to build phonetic dictionaries and alignment pipelines. While explicit language models still provide performance advantages, neural networks can implicitly learn spelling and language structure directly from training transcripts when provided with sufficient data.

Organizations developing speech technology should explore end-to-end sequence models to simplify their deployment pipelines, particularly where building phonetic lexicons is costly. For maximum transcription accuracy, hybridizing end-to-end models with existing baseline systems is recommended. Future research should focus on applying this architecture to larger datasets, spontaneous speech domains, and integrating language models directly into training.

The main limitation identified is the system's dependency on training data volume; performance was noticeably degraded on the smaller 14-hour dataset because the model requires substantial data to learn spelling and acoustic mapping simultaneously. Readers should note that while confidence is high for standard English corpora, further validation is required for highly specialized vocabularies and out-of-vocabulary terms.

Google University of Toronto audio audio data classification end-to-end speech recognition recurrent neural networks Connectionist Temporal Classification

Graves et al (2014).pdf

Paper: Speech Recognition with Deep Recurrent Neural Networks, Alex Graves et al. (2013). This paper establishes the foundational deep bidirectional LSTM architecture and Connectionist Temporal Classification training methodology that the source paper directly builds upon for end-to-end speech recognition.

Paper: Listen, Attend and Spell, William Chan et al. (2015). This work directly extends the source by replacing Connectionist Temporal Classification with an attention-based encoder-decoder architecture that eliminates conditional independence assumptions in speech recognition.
Paper: Deep Speech 2: End-to-End Speech Recognition in English and Mandarin, Dario Amodei et al. (2016). This study scales up the source paper's end-to-end Connectionist Temporal Classification approach to massive multi-thousand-hour datasets and production-grade hardware implementations.
Paper: Conformer: Convolution-augmented Transformer for Speech Recognition, Anmol Gulati et al. (2020). This paper advances the source's speech recognition framework by introducing a hybrid architecture that combines self-attention and convolutions to better capture both local and global audio patterns.

Cover for Towards End-To-End Speech Recognition with Recurrent Neural Networks

Abstract

This paper presents a speech recognition system that directly transcribes audio data with text, without requiring an intermediate phonetic representation. The system is based on a combination of the deep bidirectional LSTM recurrent neural network architecture and the Connectionist Temporal Classification objective function. A modification to the objective function is introduced that trains the network to minimise the expectation of an arbitrary transcription loss function. This allows a direct optimisation of the word error rate, even in the absence of a lexicon or language model. The system achieves a word error rate of 27.3% on the Wall Street Journal corpus with no prior linguistic information, 21.9% with only a lexicon of allowed words, and 8.2% with a trigram language model. Combining the network with a baseline system further reduces the error rate to 6.7%.

Topics

audio audio data classification long short term memory neural networks network architecture recurrent neural networks word error rate speech recognition systems optimization feedforward neural networks

Keywords

end-to-end speech recognition recurrent neural networks Connectionist Temporal Classification bidirectional LSTM bidirectional RNNs language models acoustic model stochastic gradient descent Wall Street Journal corpus

1. Introduction
2. Network Architecture
3. Connectionist Temporal Classification
4. Expected Transcription Loss
5. Decoding
6. Experiments
7. Discussion
8. Conclusion
Acknowledgements
References

Knowls

Knowl 1 — Expected Transcription Loss Objective for Direct Word Error Rate Minimization
model/method
The expected transcription loss trains a Connectionist Temporal Classification (CTC) network to directly minimize the mathematical expectation of an arbitrary sequence-level loss metric $\mathcal{L}(x, y)$ (such as the Word Error Rate) over the distribution of predicted transcriptions $y$ , given an input acoustic sequence $x = (x_1, \dots, x_T)$ of length $T$ :

$\mathcal{L}(x) = \sum_y \Pr(y|x) \mathcal{L}(x, y) = \sum_a \Pr(a|x) \mathcal{L}(x, \mathcal{B}(a))$

where $a = (a_1, \dots, a_T)$ is a frame-level alignment containing transcription labels and the blank symbol, $\mathcal{B}$ is the CTC collapse operator that removes consecutive duplicate labels and blank tokens, and $\Pr(a|x) = \prod_{t=1}^T \Pr(a_t, t|x)$ is the alignment probability given normalized softmax emission probabilities $\Pr(k, t|x) = \frac{\exp(y_t^k)}{\sum_{k'} \exp(y_t^{k'})}$ with unnormalized network outputs $y_t^k$ .

Because exact evaluation over all possible transcriptions is intractable, the expected loss and its gradient are approximated by Monte Carlo sampling. Drawing $N$ independent alignment samples $a^1, \dots, a^N \sim \Pr(a|x)$ and reusing them across all label perturbations yields a low-variance gradient estimate with respect to the unnormalized network activations $y_t^k$ :

$\frac{\partial \mathcal{L}(x)}{\partial y_t^k} \approx \frac{\Pr(k, t|x)}{N} \sum_{i=1}^N \left( \mathcal{L}(x, \mathcal{B}(a^{i,t,k})) - \mathcal{Z}(a^i, t) \right)$

where $\mathcal{Z}(a^i, t)$ is the baseline expected loss at time-step $t$ :

$\mathcal{Z}(a^i, t) = \sum_{k'} \Pr(k', t|x) \mathcal{L}(x, \mathcal{B}(a^{i,t,k'}))$

and $a^{i,t,k'}$ denotes the alignment sample $a^i$ with the label at time $t$ replaced by index $k'$ (such that $a^{i,t,k'}_{t'} = a^i_{t'}$ for all $t' \neq t$ and $a^{i,t,k'}_t = k'$ ). Reusing samples cancels baseline variance and ensures that network parameters only receive gradients for alignment modifications that alter the transcription loss.
Knowl 2 — CTC Beam Search Algorithm with Language Model and Lexicon Integration
algorithm
Decoding a Connectionist Temporal Classification (CTC) network involves finding the most probable output sequence while accounting for collapsed label repetitions, blank tokens, and optional language model constraints. The prefix beam search tracks the blank probability $\Pr^-(y, t)$ and non-blank probability $\Pr^+(y, t)$ for candidate prefixes $y$ at each time step $t \in \{1, \dots, T\}$ .

The extension probability of prefix $y$ by label $k$ at time $t$ is defined as:

$\Pr(k, y, t) = \Pr(k, t|x) \Pr(k|y) \begin{cases} \Pr^-(y, t-1) & \text{if } y^e = k \\ \Pr(y, t-1) & \text{otherwise} \end{cases}$

where $\Pr(k, t|x)$ is the CTC emission probability of label $k$ at frame $t$ , $\Pr(k|y)$ is the linguistic transition probability from prefix $y$ to $y+k$ , $y^e$ is the final label of $y$ , and $\Pr(y, t) = \Pr^-(y, t) + \Pr^+(y, t)$ is the total prefix probability with $\Pr^+( \emptyset, t) = 0$ for the empty sequence $\emptyset$ .
Input: Acoustic sequence $x$ , beam width $W$ , label set size $K$ , sequence length $T$ , transition probabilities $\Pr(k|y)$
Output: Highest probability decoded transcription sequence $y^*$
Initialize: $B \leftarrow \{\emptyset\}$ , $\Pr^-(\emptyset, 0) \leftarrow 1$ , $\Pr^+(\emptyset, 0) \leftarrow 0$
for $t = 1 \dots T$ do
     $\hat{B} \leftarrow$ the $W$ most probable sequences in $B$ based on $\Pr(y, t-1)$
     $B \leftarrow \{\}$
    for $y \in \hat{B}$ do
        if $y \neq \emptyset$ then
             $\Pr^+(y, t) \leftarrow \Pr^+(y, t-1) \Pr(y^e, t|x)$
             $\hat{y} \leftarrow$ prefix of $y$ with last label removed
            if $\hat{y} \in \hat{B}$ then
                 $\Pr^+(y, t) \leftarrow \Pr^+(y, t) + \Pr(y^e, \hat{y}, t)$
            end if
        end if
         $\Pr^-(y, t) \leftarrow \Pr(y, t-1) \Pr(-, t|x)$
        Add $y$ to $B$
        for $k = 1 \dots K$ do
             $\Pr^-(y + k, t) \leftarrow 0$
             $\Pr^+(y + k, t) \leftarrow \Pr(k, y, t)$
            Add $(y + k)$ to $B$
        end for
    end for
end for
return $\operatorname{argmax}_{y \in B} \left( \Pr(y, T) \right)^{1 / |y|}$
The length normalization factor $1 / |y|$ in the final return prevents beam search from unduly penalizing longer sequence hypotheses when decoding with an external language model.
Knowl 3 — Mapping Word-Level Language Models to Character Prefix Transition Probabilities
equation
To integrate word-level statistical language models into character-level CTC decoding, word transition distributions are mapped to character-level extension probabilities $\Pr(k|y)$ . Any partial character sequence $y$ is decomposed uniquely as $y = (w + p)$ , where $w$ is the longest complete sequence of valid dictionary words in $y$ , and $p$ is the trailing prefix of the uncompleted word (both $w$ and $p$ can be empty strings). The transition probability of extending prefix $y$ with character $k$ is given by:

$\Pr(k|y) = \frac{\sum_{w' \in (p+k)^*} \Pr^\gamma(w'|w)}{\sum_{w' \in p^*} \Pr^\gamma(w'|w)}$

where $p^*$ is the set of dictionary words prefixed by string $p$ , $(p+k)^*$ is the set of dictionary words prefixed by the concatenated string $p+k$ , $\Pr(w'|w)$ is the probability assigned by the word n-gram language model to word $w'$ given word history $w$ , and $\gamma$ is the language model scaling exponent. When using a dictionary constraint without a statistical language model, $\Pr(k|y) = 1$ if $(p+k)^*$ is non-empty and $0$ otherwise.
Knowl 4 — Deep Bidirectional LSTM-CTC Architecture for Character-Level Speech Recognition
model/method
The speech recognition network transcribes speech directly from spectral representations to character sequences without intermediate phonetic states, pronunciation lexicons, or forced frame alignments.
1. Input representation: Audio waveforms are transformed into spectrograms using Fourier transforms with 254-sample window width and 127-sample frame shifts, yielding 128 spectral coefficients per input frame.
2. Recurrent acoustic model: The architecture consists of a 5-layer deep bidirectional Long Short-Term Memory (BLSTM) network. Each hidden layer contains 500 forward LSTM cells and 500 backward LSTM cells, totaling approximately $26.5 \times 10^6$ parameters.
3. Output layer: A softmax layer contains 43 output units representing 26 uppercase English letters, punctuation marks, an explicit word delimiter space character, and the CTC blank token.
4. Training: The network is optimized end-to-end with Connectionist Temporal Classification (CTC) directly on the target character strings, integrating analytically over all valid frame-to-character alignments.
Knowl 5 — Two-Stage Training Strategy for Expected Transcription Loss Minimization
model/method
Direct optimization of the expected transcription loss (such as Word Error Rate) cannot be trained starting from random parameter initialization. For a randomly initialized network, sampled alignments almost universally produce completely erroneous transcriptions with uniformly high loss. Perturbing a single time-step label $a_t^k$ in such alignments rarely changes the sequence edit distance, causing stochastic gradient estimates to have vanishing magnitude or destructive variance.

To ensure reliable convergence, training requires a two-stage protocol:
1. Pre-training: The deep bidirectional LSTM is first trained to convergence using the standard CTC negative log-likelihood objective function $-\log \Pr(y^*|x)$ on ground-truth transcriptions $y^*$ .
2. Fine-tuning: The pre-trained network is subsequently fine-tuned using Monte Carlo expected transcription loss minimization, where drawing $N = 5$ alignment samples per utterance provides sufficiently low-variance gradient estimates.

Knowl 6 — Wall Street Journal Transcription Performance Across Language Models

data/table

The table compares Word Error Rate (WER) and Character Error Rate (CER), formatted as WER/CER where CER is available, on the Wall Street Journal (WSJ) evaluation set across different training durations (14-hour train-si84 vs. 81-hour full set) and language modeling constraints for the standard CTC model (RNN-CTC), the expected WER-trained model (RNN-WER), the baseline DNN-HMM hybrid system, and an ensemble combination:

SYSTEM	LM	14 HR	81 HR
RNN-CTC	NONE	74.2 / 30.9	30.1 / 9.2
RNN-CTC	DICTIONARY	69.2 / 30.0	24.0 / 8.0
RNN-CTC	MONOGRAM	25.8	15.8
RNN-CTC	BIGRAM	15.5	10.4
RNN-CTC	TRIGRAM	13.5	8.7
RNN-WER	NONE	74.5 / 31.3	27.3 / 8.4
RNN-WER	DICTIONARY	69.7 / 31.0	21.9 / 7.3
RNN-WER	MONOGRAM	26.0	15.2
RNN-WER	BIGRAM	15.3	9.8
RNN-WER	TRIGRAM	13.5	8.2
BASELINE	NONE	—	—
BASELINE	DICTIONARY	56.1	51.1
BASELINE	MONOGRAM	23.4	19.9
BASELINE	BIGRAM	11.6	9.4
BASELINE	TRIGRAM	9.4	7.8
COMBINATION	TRIGRAM	—	6.7

On the 81-hour dataset without a language model, the end-to-end RNN-WER achieves 27.3% WER (and 21.9% with dictionary only), substantially outperforming the baseline DNN-HMM with dictionary (51.1%). When adding a trigram language model, RNN-WER reaches 8.2% WER compared to 7.8% for the DNN-HMM baseline. When scaling from 14 to 81 hours of training data, the RNN error rate drops steeply (from 74.5% to 27.3% without LM), whereas the baseline improves only incrementally, showing that character-level models require sufficient audio data to learn implicit spelling and language regularities.

Knowl 7 — Acoustic Model Combination via N-Best Hypothesis Rescoring
model/method
The character-level RNN acoustic model and the frame-level DNN-HMM hybrid model can be combined via N-best list rescoring to achieve improved recognition accuracy.
1. Candidate generation: A list of up to $N = 300$ candidate word sequences is extracted for each utterance using the baseline DNN-HMM system with a trigram language model.
2. Candidate rescoring: Each candidate transcript $y$ is scored by computing its CTC log-probability $\log \Pr(y|x)$ under the deep bidirectional LSTM network (trained with expected WER loss).
3. Hypothesis reranking: The final score for each candidate is computed as a weighted linear combination of three components:
$\text{Score}(y) = w_{\text{DNN}} \cdot \text{Score}_{\text{DNN}}(y) + w_{\text{RNN}} \cdot \log \Pr(y|x) + w_{\text{LM}} \cdot \log \Pr_{\text{LM}}(y)$

Using weights $w_{\text{DNN}} = 1.0$ , $w_{\text{RNN}} = 1.0$ , and $w_{\text{LM}} = 11.0$ , this system combination achieves a WER of 6.7% on the 81-hour Wall Street Journal evaluation set, an absolute reduction of 1.1% over the standalone DNN-HMM baseline (7.8%) and 1.5% over the standalone RNN (8.2%).
Knowl 8 — Error Patterns and Subword Emission Dynamics in End-to-End Character ASR
empirical result
Direct character-level transcription without a dictionary exhibits specific linguistic and acoustic behaviors:
1. Error types: Unlike phonetic ASR systems that produce only in-vocabulary phoneme substitutions, the unconstrained character RNN produces three distinct categories of errors: phonetic misspellings (e.g., transcribing 'boutique' as 'bootik', or 'illustrate' as 'alstrait'), out-of-vocabulary proper noun misspellings (e.g., 'Milan' as 'Mulong', 'Dukakis' as 'Dekakis'), and homophone confusions (e.g., 'to' vs. 'two').
2. Subword acoustic grouping: Frame-level probability activations from the CTC output layer demonstrate that frequently occurring multi-character sequences (such as 'IS', 'RI', and 'END') are emitted in tightly packed bursts across adjacent frames, showing that the recurrent network learns to represent multi-character orthographic units as single acoustic sound segments.

Coverage note — None omitted; all primary contributions, including the expected transcription loss objective, the CTC beam search with language model integration, the deep BLSTM architecture, the WSJ empirical evaluation, model ensembling, and output error analyses, have been captured.

References

1.Bahl, L., Brown, P., De Souza, P.V., and Mercer, R. Maximum mutual information estimation of hidden markov model parameters for speech recognition. In Acoustics, Speech, and Signal Processing, IEEE International Conference on ICASSP ’86., volume 11, pp. 49–52, Apr 1986. doi: 10.1109/ICASSP.1986.1169179.
2.Bisani, Maximilian and Ney, Hermann. Open vocabulary speech recognition with flat hybrid models. In INTERSPEECH, pp. 725–728, 2005.
3.Bourlard, Herve A. and Morgan, Nelson. Connectionist Speech Recognition: A Hybrid Approach. Kluwer Academic Publishers, Norwell, MA, USA, 1993. ISBN 0792393961.
4.Ciresan, Dan C., Meier, Ueli, Masci, Jonathan, and Schmidhuber, Jrgen. A committee of neural networks for traffic sign classification. In IJCNN, pp. 1918–1921. IEEE, 2011.
5.Davis, S. and Mermelstein, P. Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. IEEE Transactions on Acoustics, Speech and Signal Processing, 28(4):357–366, August 1980.
6.Eyben, F., Wllmer, M., Schuller, B., and Graves, A. From speech to letters - using a novel neural network architecture for grapheme based asr. In Proc. Automatic Speech Recognition and Understanding Workshop (ASRU 2009), Merano, Italy. IEEE, 2009. 13.-17.12.2009.
7.Galescu, Lucian. Recognition of out-of-vocabulary words with sub-lexical language models. In INTERSPEECH, 2003.
8.Gers, F., Schraudolph, N., and Schmidhuber, J. Learning Precise Timing with LSTM Recurrent Networks. Journal of Machine Learning Research, 3:115–143, 2002.
9.Graves, A. and Schmidhuber, J. Framewise Phoneme Classification with Bidirectional LSTM and Other Neural Network Architectures. Neural Networks, 18(5-6):602–610, June/July 2005.
10.Graves, A., Fernández, S., Gomez, F., and Schmidhuber, J. Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks. In ICML, Pittsburgh, USA, 2006.
11.Graves, A., Mohamed, A., and Hinton, G. Speech recognition with deep recurrent neural networks. In Proc ICASSP 2013, Vancouver, Canada, May 2013.
12.Graves, Alex. Supervised Sequence Labelling with Recurrent Neural Networks, volume 385 of Studies in Computational Intelligence. Springer, 2012.
13.Hinton, G. E. and Salakhutdinov, R. R. Reducing the Dimensionality of Data with Neural Networks. Science, 313(5786):504–507, July 2006.
14.Hinton, Geoffrey, Deng, Li, Yu, Dong, Dahl, George, rahman Mohamed, Abdel, Jaitly, Navdeep, Senior, Andrew, Vanhoucke, Vincent, Nguyen, Patrick, Sainath, Tara, and Kingsbury, Brian. Deep neural networks for acoustic modeling in speech recognition. Signal Processing Magazine, 2012.
15.Hochreiter, S. and Schmidhuber, J. Long Short-Term Memory. Neural Computation, 9(8):1735–1780, 1997.
16.Jaitly, Navdeep and Hinton, Geoffrey E. Learning a better representation of speech soundwaves using restricted boltzmann machines. In ICASSP, pp. 5884–5887, 2011.
17.Jaitly, Navdeep, Nguyen, Patrick, Senior, Andrew W, and Vanhoucke, Vincent. Application of pretrained deep neural networks to large vocabulary speech recognition. In INTERSPEECH, 2012.
18.Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, 2012.
19.Lee, Li and Rose, R. A frequency warping approach to speaker normalization. Speech and Audio Processing, IEEE Transactions on, 6(1):49–60, Jan 1998.
20.Peters, J. and Schaal, S. Reinforcement learning of motor skills with policy gradients. In Neural Networks, number 4, pp. 682–97, 2008.
21.Povey, D., Ghoshal, A., Boulianne, G., Burget, L., Glembek, O., Goel, N., Hannemann, M., Motlicek, P., Qian, Y., Schwarz, P., Silovsky, J., Stemmer, G., and Vesely, K. The kaldi speech recognition toolkit. In IEEE 2011 Workshop on Automatic Speech Recognition and Understanding. IEEE Signal Processing Society, December 2011.
22.Schuster, M. and Paliwal, K. K. Bidirectional Recurrent Neural Networks. IEEE Transactions on Signal Processing, 45:2673–2681, 1997.

Citation

MLA

Graves, A., and N. Jaitly. “Towards End-To-End Speech Recognition with Recurrent Neural Networks”. International Conference on Machine Learning, 2014, pp. 1764–72, http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.664.3860.

APA

Graves, A., & Jaitly, N. (2014). Towards End-To-End Speech Recognition with Recurrent Neural Networks. International Conference on Machine Learning, 1764–1772. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.664.3860

Chicago

Graves, A., and N. Jaitly. 2014. “Towards End-To-End Speech Recognition with Recurrent Neural Networks”. International Conference on Machine Learning, 1764–72. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.664.3860.

Harvard

Graves, A. and Jaitly, N. (2014) “Towards End-To-End Speech Recognition with Recurrent Neural Networks”, International Conference on Machine Learning, pp. 1764–1772. Available at: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.664.3860.

Vancouver

1. Graves A, Jaitly N (2014) Towards End-To-End Speech Recognition with Recurrent Neural Networks. International Conference on Machine Learning 1764–1772

BibTeX

@article{graves2014towards,
  title = {Towards End-To-End Speech Recognition with Recurrent Neural Networks},
  author = {Graves, Alex and Jaitly, Navdeep},
  year = {2014},
  journal = {International Conference on Machine Learning},
  pages = {1764-1772},
  url = {http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.664.3860}
}

Metadata:DOI registry

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Published with permission

Towards End-To-End Speech Recognition with Recurrent Neural Networks

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Expected Transcription Loss Objective for Direct Word Error Rate Minimization

Knowl 2 — CTC Beam Search Algorithm with Language Model and Lexicon Integration

Knowl 3 — Mapping Word-Level Language Models to Character Prefix Transition Probabilities

Knowl 4 — Deep Bidirectional LSTM-CTC Architecture for Character-Level Speech Recognition

Knowl 5 — Two-Stage Training Strategy for Expected Transcription Loss Minimization

Knowl 6 — Wall Street Journal Transcription Performance Across Language Models

Knowl 7 — Acoustic Model Combination via N-Best Hypothesis Rescoring

Knowl 8 — Error Patterns and Subword Emission Dynamics in End-to-End Character ASR

References

Citation

BibTeX

Access the Paper