*DeepSpeech2 on PaddlePaddle* is an open-source implementation of end-to-end Automatic Speech Recognition (ASR) engine, based on [Baidu's Deep Speech 2 paper](http://proceedings.mlr.press/v48/amodei16.pdf), with [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform. Our vision is to empower both industrial application and academic research on speech-to-text, via an easy-to-use, efficent and scalable integreted implementation, including training & inferencing module, distributed [PaddleCloud](https://github.com/PaddlePaddle/cloud) training, and demo deployment. Besides, several pre-trained models for both English and Mandarin speech are also released.
*DeepSpeech2 on PaddlePaddle* is an open-source implementation of end-to-end Automatic Speech Recognition (ASR) engine, based on [Baidu's Deep Speech 2 paper](http://proceedings.mlr.press/v48/amodei16.pdf), with [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) platform. Our vision is to empower both industrial application and academic research on speech-to-text, via an easy-to-use, efficent and scalable integreted implementation, including training, inferencing & testing module, distributed [PaddleCloud](https://github.com/PaddlePaddle/cloud) training, and demo deployment. Besides, several pre-trained models for both English and Mandarin are also released.
Several shell scripts provided in `./examples` will help us to quickly give it a try, including training, inferencing, evaluation and demo deployment.
Several shell scripts provided in `./examples` will help us to quickly give it a try, for most major modules, including data preparation, model training, case inference, model evaluation and demo deployment, with a few public dataset (e.g. [LibriSpeech](http://www.openslr.org/12/), [Aishell](https://github.com/kaldi-asr/kaldi/tree/master/egs/aishell)). Reading these examples will also help us understand how to make it work with our own data.
Most of the scripts in `./examples` are configured with 8 GPUs. If you don't have 8 GPUs available, please modify `CUDA_VISIBLE_DEVICE` and `--trainer_count`. If you don't have any GPU available, please set `--use_gpu` to False.
Some of the scripts in `./examples` are configured with 8 GPUs. If you don't have 8 GPUs available, please modify `CUDA_VISIBLE_DEVICE` and `--trainer_count`. If you don't have any GPU available, please set `--use_gpu` to False to use CPUs instead.
Let's take a tiny sampled subset of [LibriSpeech dataset](http://www.openslr.org/12/) for instance.
Let's take a tiny sampled subset of [LibriSpeech dataset](http://www.openslr.org/12/) for instance.
- Go to directory
- Go to directory
```
```
cd examples/librispeech_tiny
cd examples/tiny
```
```
Notice that this is only a toy example with a tiny sampled set of LibriSpeech. If we would like to try with the complete LibriSpeech (would take much a longer time for training), please go to `examples/librispeech` instead.
Notice that this is only a toy example with a tiny sampled subset of LibriSpeech. If we would like to try with the complete dataset (would take several days for training), please go to `examples/librispeech` instead.
- Prepare the libripseech data
- Prepare the data
```
```
sh preprare_data.sh
sh run_data.sh
```
```
`prepare_data.sh` downloads dataset, generates file manifests, collects normalizer' statitics and builds vocabulary for us. Once the running is done, we'll find our LibriSpeech data (not full in this "tiny" example) downloaded in `~/.cache/paddle/dataset/speech/Libri` and several manifest files as well as one mean stddev file generated in `./data/librispeech_tiny`, for the further model training. It needs to be run for only once.
`run_data.sh` will download dataset, generate manifests, collect normalizer' statitics and build vocabulary. Once the data preparation is done, we will find the data (only part of LibriSpeech) downloaded in `~/.cache/paddle/dataset/speech/libri` and the corresponding manifest files generated in `./data/tiny` as well as a mean stddev file and a vocabulary file. It has to be run for the very first time we run this dataset and is reusable for all further experiments.
- Train your own ASR model
- Train your own ASR model
```
```
sh run_train.sh
sh run_train.sh
```
```
`run_train.sh` starts a training job, with training logs printed to stdout and model checkpoint of every pass/epoch saved to `./checkpoints`. We can resume the training from these checkpoints, or use them for inference, evalutiaton and deployment.
`run_train.sh` will start a training job, with training logs printed to stdout and model checkpoint of every pass/epoch saved to `./checkpoints/tiny`. We can resume the training from these checkpoints, or use them for inference, evalutiaton and deployment.
- Case inference with an existing model
- Case inference with an existing model
```
```
sh run_infer.sh
sh run_infer.sh
```
```
`run_infer.sh` will quickly show us speech-to-text decoding results for several (default: 10) audio samples with an existing model. Since the model is only trained on a subset of LibriSpeech, the performance might not be very good. We can download a well-trained model and then do the inference:
`run_infer.sh` will show us some speech-to-text decoding results for several (default: 10) samples with the trained model. The performance might not be good now as the current model is only trained with a toy subset of LibriSpeech. To see the results with a better model, we can download a well-trained (trained for several days, with the complete LibriSpeech) model and do the inference:
```
```
sh download_model_run_infer.sh
sh run_infer_golden.sh
```
```
- Evaluate an existing model
- Evaluate an existing model
...
@@ -76,14 +78,14 @@ Let's take a tiny sampled subset of [LibriSpeech dataset](http://www.openslr.org
...
@@ -76,14 +78,14 @@ Let's take a tiny sampled subset of [LibriSpeech dataset](http://www.openslr.org
sh run_test.sh
sh run_test.sh
```
```
`run_test.sh` evaluates the model with Word Error Rate (or Character Error Rate) measurement. Similarly, we can also download a well-trained model and test its performance:
`run_test.sh` will evaluate the model with Word Error Rate (or Character Error Rate) measurement. Similarly, we can also download a well-trained model and test its performance:
```
```
sh download_model_run_test.sh
sh run_test_golden.sh
```
```
- Try out a live demo with your own voice
- Try out a live demo with your own voice
Until now, we have trained and tested an ASR model quantitively and qualitatively with existing audios. But we haven't try the model with our own speech. `demo_server.sh` and `demo_client.sh` helps quickly build up a demo ASR engine with the trained model, enabling us to test and play around with the demo with our own voice.
Until now, we have trained and tested our ASR model qualitatively (`run_infer.sh`) and quantitively (`run_test.sh`) with existing audio files. But we have not yet play the model with our own speech. `demo_server.sh` and `demo_client.sh` helps quickly build up a demo ASR engine with the trained model, enabling us to test and play around with the demo with our own voice.
We start the server in one console by entering:
We start the server in one console by entering:
...
@@ -112,20 +114,20 @@ Wish you a happy journey with the DeepSpeech2 ASR engine!
...
@@ -112,20 +114,20 @@ Wish you a happy journey with the DeepSpeech2 ASR engine!
#### Generate Manifest
#### Generate Manifest
*DeepSpeech2 on PaddlePaddle* accepts a textual **manifest** file as its data set interface. A manifest file summarizes a set of speech data, with each line containing the meta data (e.g. filepath, transcription, duration) of one audio clip, in [JSON](http://www.json.org/) format, just as:
*DeepSpeech2 on PaddlePaddle* accepts a textual **manifest** file as its data set interface. A manifest file summarizes a set of speech data, with each line containing some meta data (e.g. filepath, transcription, duration) of one audio clip, in [JSON](http://www.json.org/) format, such as:
```
```
{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0001.flac", "duration": 3.275, "text": "stuff it into you his belly counselled him"}
{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0001.flac", "duration": 3.275, "text": "stuff it into you his belly counselled him"}
{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0007.flac", "duration": 4.275, "text": "a cold lucid indifference reigned in his soul"}
{"audio_filepath": "/home/work/.cache/paddle/Libri/134686/1089-134686-0007.flac", "duration": 4.275, "text": "a cold lucid indifference reigned in his soul"}
```
```
To use any custom data, we only need to generate such manifest files to summarize the dataset. Given such summarized manifests, training, inference and all other modules can be aware of where to access the audio files, as well as their meta data including the transcription labels.
To use your custom data, you only need to generate such manifest files to summarize the dataset. Given such summarized manifests, training, inference and all other modules can be aware of where to access the audio files, as well as their meta data including the transcription labels.
For example script to generate such manifest files, please refer to `data/librispeech/librispeech.py`, which download and generate manifests for LibriSpeech dataset.
For how to generate such manifest files, please refer to `data/librispeech/librispeech.py`, which download and generate manifests for LibriSpeech dataset.
#### Compute Mean & Stddev for Normalizer
#### Compute Mean & Stddev for Normalizer
To perform z-score normalization (zero-mean, unit stddev) upon audio features, we have to estimate in advance the mean and standard deviation of the features, with sampled training audios:
To perform z-score normalization (zero-mean, unit stddev) upon audio features, we have to estimate in advance the mean and standard deviation of the features, with some training samples:
```
```
python tools/compute_mean_std.py \
python tools/compute_mean_std.py \
...
@@ -140,7 +142,7 @@ It will compute the mean and standard deviation of power spectgram feature with
...
@@ -140,7 +142,7 @@ It will compute the mean and standard deviation of power spectgram feature with
#### Build Vocabulary
#### Build Vocabulary
A list of possible characters is required to convert the target transcription into list of token indices for training and in docoders convert from them back to text. Such a character-based vocabulary can be build with `tools/build_vocab.py`.
A vocabulary of possible characters is required to convert the transcription into a list of token indices for training, and in docoding, to convert from a list of indices back to text again. Such a character-based vocabulary can be build with `tools/build_vocab.py`.
```
```
python tools/build_vocab.py \
python tools/build_vocab.py \
...
@@ -149,7 +151,7 @@ python tools/build_vocab.py \
...
@@ -149,7 +151,7 @@ python tools/build_vocab.py \
--manifest_paths data/librispeech/manifest.train
--manifest_paths data/librispeech/manifest.train
```
```
It will build a vocabuary file of `data/librispeeech/eng_vocab.txt` with all transcription text in `data/librispeech/manifest.train`, without character truncation.
It will write a vocabuary file `data/librispeeech/eng_vocab.txt` with all transcription text in `data/librispeech/manifest.train`, without vocabulary truncation (`--count_threshold 0`).
Data augmentation has often been a highly effective technique to boost the deep learning performance. We augment our speech data by synthesizing new audios with small random perterbation (label-invariant transformation) added upon raw audios. We don't have to do the syntheses by ourselves, as it is already embeded into the data provider and is done on the fly, randomly for each epoch.
Data augmentation has often been a highly effective technique to boost the deep learning performance. We augment our speech data by synthesizing new audios with small random perterbation (label-invariant transformation) added upon raw audios. We don't have to do the syntheses by ourselves, as it is already embeded into the data provider and is done on the fly, randomly for each epoch during training.
Six optional augmentation components are provided for us to configured and inserted into the processing pipeline.
Six optional augmentation components are provided for us to configured and inserted into the processing pipeline.
...
@@ -203,7 +205,7 @@ Six optional augmentation components are provided for us to configured and inser
...
@@ -203,7 +205,7 @@ Six optional augmentation components are provided for us to configured and inser
In order to inform the trainer of what augmentation components we need and what their processing orders are, we are required to prepare a *augmentation configuration file* in JSON format. For example:
In order to inform the trainer of what augmentation components we need and what their processing orders are, we are required to prepare a *augmentation configuration file* in [JSON](http://www.json.org/) format. For example:
```
```
[{
[{
...
@@ -220,23 +222,23 @@ In order to inform the trainer of what augmentation components we need and what
...
@@ -220,23 +222,23 @@ In order to inform the trainer of what augmentation components we need and what
}]
}]
```
```
When the `--augment_conf_file` argument of `trainer.py` is set to the path of the above example configuration file, each audio clip in each epoch will be processed: with 60% of chance, it will first be speed perturbed with a uniformly random sampled speed-rate between 0.95 and 1.05, and then with 80% of chance it will be shifted in time with a random sampled offset between -5 ms and 5 ms. Finally this newly synthesized audio clip will be feed into the feature extractor for further training.
When the `--augment_conf_file` argument of `trainer.py` is set to the path of the above example configuration file, every audio clip in every epoch will be processed: with 60% of chance, it will first be speed perturbed with a uniformly random sampled speed-rate between 0.95 and 1.05, and then with 80% of chance it will be shifted in time with a random sampled offset between -5 ms and 5 ms. Finally this newly synthesized audio clip will be feed into the feature extractor for further training.
For configuration examples, please refer to `conf/augmenatation.config.example`.
For other configuration examples, please refer to `conf/augmenatation.config.example`.
Be careful when we are utilizing the data augmentation technique, as improper augmentation will instead do harm to the training, due to the enlarged train-test gap.
Be careful when we are utilizing the data augmentation technique, as improper augmentation will do harm to the training, due to the enlarged train-test gap.
## Inference and Evaluation
## Inference and Evaluation
#### Prepare Language Model
### Prepare Language Model
A language model is required to improve the decoder's performance. We have prepared two language models (with lossy compression) for users to download and try. One is for English and the other is for Mandarin. Please refer to `examples/librispeech/download_model.sh` and `examples/mandarin_demo/download_model.sh` for their urls. If you wish to train your own better language model, please refer to [KenLM](https://github.com/kpu/kenlm) for tutorials.
A language model is required to improve the decoder's performance. We have prepared two language models (with lossy compression) for users to download and try. One is for English and the other is for Mandarin. Please refer to `models/lm/download_lm_en.sh` and `models/lm/download_lm_zh.sh` for their urls. If you wish to train your own better language model, please refer to [KenLM](https://github.com/kpu/kenlm) for tutorials.
TODO: any other requirements or tips to add?
TODO: any other requirements or tips to add?
#### Speech-to-text Inference
### Speech-to-text Inference
We provide a inference module `infer.py` to infer, decode and visualize speech-to-text results for several given audio clips, which might help to have a intuitive and qualitative evaluation of the ASR model performance.
We provide a inference module `infer.py` to infer, decode and visualize speech-to-text results for several given audio clips. It might help us to have a intuitive and qualitative evaluation of the ASR model's performance.
- Inference with GPU:
- Inference with GPU:
...
@@ -247,21 +249,21 @@ We provide a inference module `infer.py` to infer, decode and visualize speech-t
...
@@ -247,21 +249,21 @@ We provide a inference module `infer.py` to infer, decode and visualize speech-t
We provide two CTC decoders: *CTC greedy decoder* and *CTC beam search decoder*. The *CTC greedy decoder* is an implementation of the simple best-path decoding algorithm, selecting at each timestep the most likely token, thus being greedy and locally optimal. The [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) otherwise utilzied a heuristic breadth-first gragh search for arriving at a near global optimality; it requires a pre-trained KenLM language model for better scoring and ranking sentences. The decoder type can be set with argument `--decoding_method`.
We provide two types of CTC decoders: *CTC greedy decoder* and *CTC beam search decoder*. The *CTC greedy decoder* is an implementation of the simple best-path decoding algorithm, selecting at each timestep the most likely token, thus being greedy and locally optimal. The [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) otherwise utilizes a heuristic breadth-first gragh search for reaching a near global optimality; it also requires a pre-trained KenLM language model for better scoring and ranking. The decoder type can be set with argument `--decoding_method`.
For more help on arguments:
For more help on arguments:
```
```
python infer.py --help
python infer.py --help
```
```
or refer to `example/librispeech/run_infer.sh.
or refer to `example/librispeech/run_infer.sh`.
#### Evaluate a Model
### Evaluate a Model
To evaluate a model quantitively, we can run:
To evaluate a model's performance quantitively, we can run:
- Evaluation with GPU:
- Evaluation with GPU:
...
@@ -272,23 +274,23 @@ To evaluate a model quantitively, we can run:
...
@@ -272,23 +274,23 @@ To evaluate a model quantitively, we can run:
- Evaluation with CPU:
- Evaluation with CPU:
```
```
python test.py --use_gpu False
python test.py --use_gpu False --trainer_count 12
```
```
The error rate (default: word error rate, can be set with `--error_rate_type`) will be printed.
The error rate (default: word error rate; can be set with `--error_rate_type`) will be printed.
For more help on arguments:
For more help on arguments:
```
```
python test.py --help
python test.py --help
```
```
or refer to `example/librispeech/run_test.sh.
or refer to `example/librispeech/run_test.sh`.
## Hyper-parameters Tuning
## Hyper-parameters Tuning
The hyper-parameters $\alpha$ (coefficient for language model scorer) and $\beta$ (coefficient for word count scorer) for the [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) often have a significant impact on the decoder's performance. It'd be better to re-tune them on validation samples after the accustic model is renewed.
The hyper-parameters $\alpha$ (coefficient for language model scorer) and $\beta$ (coefficient for word count scorer) for the [*CTC beam search decoder*](https://arxiv.org/abs/1408.2873) often have a significant impact on the decoder's performance. It would be better to re-tune them on a validation set when the accustic model is renewed.
`tools/tune.py` performs a 2-D grid search over the hyper-parameter $\alpha$ and $\beta$. We have to provide the range of $\alpha$ and $\beta$, as well as the number of attempts.
`tools/tune.py` performs a 2-D grid search over the hyper-parameter $\alpha$ and $\beta$. We have to provide the range of $\alpha$ and $\beta$, as well as the number of their attempts.
- Tuning with GPU:
- Tuning with GPU:
...
@@ -309,12 +311,12 @@ The hyper-parameters $\alpha$ (coefficient for language model scorer) and $\beta
...
@@ -309,12 +311,12 @@ The hyper-parameters $\alpha$ (coefficient for language model scorer) and $\beta
python tools/tune.py --use_gpu False
python tools/tune.py --use_gpu False
```
```
After tuning, we can reset $\alpha$ and $\beta$ in the inference and evaluation modules to see if they can really improve the ASR performance.
After tuning, we can reset $\alpha$ and $\beta$ in the inference and evaluation modules to see if they really help improve the ASR performance.
```
```
python tune.py --help
python tune.py --help
```
```
or refer to `example/librispeech/run_tune.sh.
or refer to `example/librispeech/run_tune.sh`.
TODO: add figure.
TODO: add figure.
...
@@ -352,4 +354,6 @@ It could be possible to start the server and the client in two seperate machines
...
@@ -352,4 +354,6 @@ It could be possible to start the server and the client in two seperate machines
