deep_speech_2.md

# DeepSpeech2 on PaddlePaddle: Design Doc 

We are planning to build Deep Speech 2 (DS2) \[[1](#references)\], a powerful Automatic Speech Recognition (ASR) engine,  on PaddlePaddle. For the first-stage plan, we have the following short-term goals:

- Release a basic distributed implementation of DS2 on PaddlePaddle.
- Contribute a chapter of Deep Speech to PaddlePaddle Book.

Intensive system optimization and low-latency inference library (details in \[[1](#references)\]) are not yet covered in this first-stage plan.

## Table of Contents

- [Tasks](#tasks)
- [Task Dependency](#task-dependency)
- [Design Details](#design-details)
    - [Overview](#overview)
    - [Row Convolution](#row-convolution)
    - [Beam Search With CTC and LM](#beam-search-with-ctc-and-lm)
- [Future Work](#future-work)
- [References](#references)

## Tasks

We roughly break down the project into 14 tasks:

1. Develop an **audio data provider**:
	- Json filelist generator.
	- Audio file format transformer.
	- Spectrogram feature extraction, power normalization etc.
	- Batch data reader with SortaGrad.
	- Data augmentation (optional).
	- Prepare (one or more) public English data sets & baseline.
2. Create a **simplified DS2 model configuration**:
   - With only fixed-length (by padding) audio sequences (otherwise need *Task 3*).
	- With only bidirectional-GRU (otherwise need *Task 4*).
	- With only greedy decoder (otherwise need *Task 5, 6*).
3. Develop to support **variable-shaped** dense-vector (image) batches of input data.
   - Update `DenseScanner` in `dataprovider_converter.py`, etc.
4. Develop a new **lookahead-row-convolution layer** (See \[[1](#references)\] for details):
   - Lookahead convolution windows.
   - Within-row convolution, without kernels shared across rows.
5. Build KenLM **language model** (5-gram) for beam search decoder:
   - Use KenLM toolkit.
   - Prepare the corpus & train the model.
   - Create infererence interfaces (for Task 6).
6. Develop a **beam search decoder** with CTC + LM + WORDCOUNT:
   - Beam search with CTC.
   - Beam search with external custom scorer (e.g. LM).
   - Try to design a more general beam search interface.
7. Develop a **Word Error Rate evaluator**:
   - update `ctc_error_evaluator`(CER) to support WER.
8. Prepare internal dataset for Mandarin (optional):
    - Dataset, baseline, evaluation details.
    - Particular data preprocessing for Mandarin.
    - Might need cooperating with the Speech Department.
9. Create **standard DS2 model configuration**:
   - With variable-length audio sequences (need *Task 3*).
	- With unidirectional-GRU + row-convolution (need *Task 4*).
	- With CTC-LM beam search decoder (need *Task 5, 6*).
10. Make it run perfectly on **clusters**.
11. Experiments and **benchmarking** (for accuracy, not efficiency):
    - With public English dataset.
    - With internal (Baidu) Mandarin dataset (optional).
12. Time **profiling** and optimization.
13. Prepare **docs**.
14. Prepare PaddlePaddle **Book** chapter with a simplified version.

## Task Dependency

Tasks parallelizable within phases:

Roadmap     | Description                               | Parallelizable Tasks 
----------- | :------------------------------------     | :--------------------
Phase I	    | Simplified model & components             | *Task 1* ~ *Task 8*
Phase II    | Standard model & benchmarking & profiling | *Task 9* ~ *Task 12*
Phase III   | Documentations                            | *Task13* ~ *Task14*

Issue for each task will be created later. Contributions, discussions and comments are all highly appreciated and welcomed!

## Design Details

### Overview

Traditional **ASR** (Automatic Speech Recognition) pipelines require great human efforts devoted to elaborately tuning multiple hand-engineered components (e.g. audio feature design, accoustic model, pronuncation model and language model etc.). **Deep Speech 2** (**DS2**) \[[1](#references)\], however, trains such ASR models in an end-to-end manner, replacing most intermediate modules with only a single deep network architecture. With scaling up both the data and model sizes, DS2 achieves a very significant performance boost.

Please read Deep Speech 2 \[[1](#references),[2](#references)\] paper for more background knowledge.

The classical DS2 network contains 15 layers (from bottom to top):

- **Two** data layers (audio spectrogram, transcription text)
- **Three** 2D convolution layers
- **Seven** uni-directional simple-RNN layers
- **One** lookahead row convolution layers
- **One** fully-connected layers
- **One** CTC-loss layer

<div align="center">
<img src="image/ds2_network.png" width=350><br/>
Figure 1. Archetecture of Deep Speech 2 Network.
</div>

We don't have to persist on this 2-3-7-1-1-1 depth \[[2](#references)\]. Similar networks with different depths might also work well. As in \[[1](#references)\], authors use a different depth (e.g. 2-2-3-1-1-1) for final experiments.

Key ingredients about the layers:

- **Data Layers**: 
   - Frame sequences data of audio **spectrogram** (with FFT).
   - Token sequences data of **transcription** text (labels). 
   - These two type of sequences do not have the same lengthes, thus a CTC-loss layer is required.
- **2D Convolution Layers**: 
   - Not only temporal convolution, but also **frequency convolution**. Like a 2D image convolution, but with a variable dimension (i.e. temporal dimension).
   - With striding for only the first convlution layer.
   - No pooling for all convolution layers.
- **Uni-directional RNNs** 
	- Uni-directional + row convolution: for low-latency inference.
	- Bi-direcitional + without row convolution: if we don't care about the inference latency.
- **Row convolution**:
	- For looking only a few steps ahead into the feature, instead of looking into a whole sequence in bi-directional RNNs.
	- Not nessesary if with bi-direcitional RNNs. 
	- "**Row**" means convolutions are done within each frequency dimension (row), and no convolution kernels shared across.
- **Batch Normalization Layers**:
   - Added to all above layers (except for data and loss layer).
   - Sequence-wise normalization for RNNs: BatchNorm only performed on input-state projection and not state-state projection, for efficiency consideration.
 

Required Components                     | PaddlePaddle Support                      | Need to Develop
:-------------------------------------  | :--------------------------------------   | :-----------------------
Data Layer I (Spectrogram)	            | Not supported yet.                        |  TBD (Task 3)
Data Layer II (Transcription)           | `paddle.data_type.integer_value_sequence` | -
2D Convolution Layer                    | `paddle.layer.image_conv_layer`           | -
DataType Converter (vec2seq)            | `paddle.layer.block_expand`               | -
Bi-/Uni-directional RNNs                | `paddle.layer.recurrent_group`            | -
Row Convolution Layer                   | Not supported yet.                        | TBD (Task 4)
CTC-loss Layer                          | `paddle.layer.warp_ctc`                   | -
Batch Normalization Layer               | `paddle.layer.batch_norm`                 | -
CTC-Beam search                         | Not supported yet.                        | TBD (Task 6)

### Row Convolution

TODO by Assignees

### Beam Search with CTC and LM

<div align="center">
<img src="image/beam_search.png" width=600><br/>
Figure 2. Algorithm for CTC Beam Search Decoder.
</div>

- The **Beam Search Decoder** for DS2 CTC-trained network follows the similar approach in \[[3](#references)\] as shown in Figure 2, with two important modifications for the ambiguous parts: 
   - 1) in the iterative computation of probabilities, the assignment operation is changed to accumulation for one prefix may comes from different paths; 
   - 2) the if condition ```if l^+ not in A_prev then``` after probabilities' computation is deprecated for it is hard to understand and seems unnecessary.
- An **external scorer** would be passed into the decoder to evaluate a candidate prefix during decoding whenever a white space appended in English decoding and any character appended in Mandarin decoding.
- Such external scorer consists of language model, word count or any other custom scorers.
- The **language model** is built from Task 5, with parameters should be carefully tuned to achieve minimum WER/CER (c.f. Task 7)
- This decoder needs to perform with **high efficiency** for the convenience of parameters tuning and speech recognition in reality. 
 

## Future Work

- Efficiency Improvement
- Accuracy Improvement
- Low-latency Inference Library
- Large-scale benchmarking

## References

1. Dario Amodei, etc., [Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin](http://proceedings.mlr.press/v48/amodei16.pdf). ICML 2016.
2. Dario Amodei, etc., [Deep Speech 2 : End-to-End Speech Recognition in English and Mandarin](https://arxiv.org/abs/1512.02595). 	arXiv:1512.02595.
3. Awni Y. Hannun, etc. [First-Pass Large Vocabulary Continuous Speech Recognition using Bi-Directional Recurrent DNNs](https://arxiv.org/abs/1408.2873). arXiv:1408.2873