Speech Spoofing

Speech Spoofing

The goal of this system if to convert the voice of one person to another person without altering the utterance.

Workflow

Steps

1. Data Collection
2. Feature Extraction
3. Mapping
4. Reconstruction

Future Extensions

1. PSOLA: This method can be used for reconstruction of speech signal effectively. It is supported in Praat Toolkit.
2. STRAIGHT: This is non open source toolkit for feature extraction and speech reconstruction maintaining the quality of speech signal.
3. Other Neural Network Architectures like Sequence to Sequence Models with Attention Framework can be used to better model the mapping function between the source and target.
4. DTW: In case the available data set is not aligned in time, this can be done using Dynamic Time Warping. It can be implemented in Python using librosa.core.dtw

Applications

1. Synthetic Speech Production.
2. Entertainment Industry (Dubbing).

Installations

1. pyrenn
2. pysptk
3. librosa
4. numpy
5. keras

References

• Voice Conversion Implementation
• Overview
• Voice Conversion using ANN
• Voice Conversion using Deep RNN
• Thesis on VC using Deep RNN