Speaker Recognition

Speaker Recognition

The built system could classify 10 voices with 100% accuracy tested on the ELSDSR dataset with 90 seconds of test speech per speaker.

Workflow

Steps

1. Data Collection
2. Feature Extraction
3. Model Building & Training
4. Classification

Future Extensions

1. Confusion Estimate.
2. Create a threshold to identify if the speaker is new.
3. User Interface to Train & Test using Tkinter.
4. Testing with Real Life / Noisy Data.

Applications

1. Attendance in Classrooms.
2. Voice Based Security in addition to Biometrics.

Installations

1. NumPy: numpy
2. Hidden Markov Models: hmmlearn
3. Speech Signal Analysis: librosa

References

• Text Independent Speaker Recognition using HMMs
• Speaker Recognition using GMM-HMM Hybrid Model
• Hidden Markov Models in Speech Applications by Rabiner