Skip to content

Latest commit

 

History

History
24 lines (15 loc) · 2.3 KB

README.md

File metadata and controls

24 lines (15 loc) · 2.3 KB

babble-rnn: Generating speech from speech with LSTM networks

babble-rnn is a research project in the use of machine learning to generate new speech by modelling human speech audio, without any intermediate text or word representations. The idea is to learn to speak through imitation, much like a baby might. The goal is to generate a babbling audio output that emulates the speech patterns of the original speaker, ideally incorporating real words into the output.

The implementation is based on Keras / Theano, generating an LSTM RNN; and Codec 2, an open source speech audio compression algorithm. The resulting models have learned the most common audio sequences of a 'performer', and can generate a probable babbling audio sequence when provided a seed sequence.

Read the babble-rnn tech post

View the babble-rnn code on Github

Wondering what babble-rnn can do? Listen to the latest babble produced by the experiments since the original tech report:

play the audio

This babbler is a stack of 11 bidirectional LSTMs, attempting to learning an encoded sequence of data (frame of 13 normalized parameters, representing 20ms of audio). Groups of LSTMs are trained together, while keeping others locked, to limit the complexity of learning such a deep network.

The audio itself is highly compressed through the Codec 2 (see the original tech post for details) producing a 3200 bit per second stream of frequency, energy, sinusoidal and voicing parameters. An autoencoder learns the features of this against a particular human speaker, to compress the output further. The encoder stage is a mix of 2D convolutional layers, picking features from the Codec 2 data over short time sequences, runnning in parallel with a series of standard hidden layers (to provide a compressed stream that helps feed through some of the original input), before being merged into a single encoded output at a quarter of the rate of the original Codec 2 input (80ms audio per frame, although more parameters output than the original).