wenlin-zhang / speecht

An opensource speech-to-text software written in tensorflow

speechT

An opensource speech-to-text software written in tensorflow.

Python 3 is required.

Architecture

Currently speechT uses a simple sequence-to-sequence model with GRU cells.
The input consists of raw audio data imported using soundfile.
The decoder uses an attention based mechanism, target words are embedded first.

The speech corpus from http://www.openslr.org/12/ is automatically downloaded.
Note: The corpus is about 30GB!

See also Sequence-to-Sequence models in tensorflow.

Training

To start training just run

python3 transcribe.py

Important flags
--data_dir to specify the data directory to download speech corpus to (defaults to ./data/)
--train_dir to specify the train directory to save checkpoints and vocabulary to (defaults to ./train/)
--size size of each model layer (defaults to 1024)
--num_layers number of layers in the model (defaults to 3)
--vocab_size maximum vocabulary size (defaults to 40000)

Testing

To test your trained model run

python3 transcribe.py --decode_tests [--data_dir PATH] [--train_dir PATH]

Live usage

To record using your microphone and convert it to text run

python3 transcribe.py --decode [--data_dir PATH] [--train_dir PATH]