superhg2012 / speaker-recognition

A Real-time Speaker Recognition System with GUI

About

This is a Speaker Recognition system with GUI.

For more details of this project, please see:

• Our presentation slides
• Our complete report

Dependencies

• scikit-learn
• scikits.talkbox
• pyssp
• PyQt4
• PyAudio
• (Optional)Python bindings for bob:
  • install blitz, openblas, boost
  • pip install --user bob.extension bob.blitz bob.core bob.sp bob.ap

Note: We have a MFCC implementation on our own which will be used as a fallback when bob is unavailable. But it's not so efficient as the C implementation in bob.

Compile GMM (Optional)

Run make -C src/gmm. Require gcc >= 4.7. If compiled successfully, it will be used by default instead of GMM from scikit-learn. But it doesn't make much difference apart from the speed.

Algorithms Used

Voice Activity Detection(VAD):

• Long-Term Spectral Divergence (LTSD)

Feature:

• Mel-Frequency Cepstral Coefficient (MFCC)
• Linear Predictive Coding (LPC)

Model:

• Gaussian Mixture Model (GMM)
• Universal Background Model (UBM)
• Continuous Restricted Boltzman Machine (CRBM)
• Joint Factor Analysis (JFA)

GUI Demo

Our GUI not only has basic functionality for recording, enrollment, training and testing, but also has a visualization of real-time speaker recognition:

You can See our demo video (in Chinese). Note that real-time speaker recognition is extremely hard, because we only use corpus of about 1 second length to identify the speaker. Therefore the real-time system doesn't work very perfect.

Also the GUI part is quite hacky for demo purpose and may not work smoothly anymore today.

Command Line Tools

usage: speaker-recognition.py [-h] -t TASK -i INPUT -m MODEL

Speaker Recognition Command Line Tool

optional arguments:
  -h, --help            show this help message and exit
  -t TASK, --task TASK  Task to do. Either "enroll" or "predict"
  -i INPUT, --input INPUT
                        Input Files(to predict) or Directories(to enroll)
  -m MODEL, --model MODEL
                        Model file to save(in enroll) or use(in predict)

Wav files in each input directory will be labeled as the basename of the directory.
Note that wildcard inputs should be *quoted*, and they will be sent to glob module.

Examples:
    Train:
    ./speaker-recognition.py -t enroll -i "./bob/ ./mary/ ./person*" -m model.out

    Predict:
    ./speaker-recognition.py -t predict -i "./*.wav" -m model.out