← ALL WORK

05 / Applied AI / 2026

Speech modelling and detection.

Speech technology / VAD + GMM

CONTRIBUTION

Individual coursework

BUILT WITH
PythonPyTorchscikit-learnGMM
View source

01 / SCOPE

Problem and
scope.

The evaluation compares neural voice-activity detectors and classical models across speech corpora, with the cross-corpus result as the main test.

MY ROLE

Individual modelling and evaluation work using supplied datasets and practical frontends. The neural VAD and GMM implementations are available in the public repository.

02 / IMPLEMENTATION

Implementation
details.

  1. 01Acoustic features
  2. 02Context / sequence
  3. 03Neural VAD
  4. 04Post-processing
  5. 05Cross-corpus test
01

Compare model families.

Move from a contextual MLP to bidirectional recurrent models and dilated temporal convolutions, examining how temporal context changes detection.

02

Test on a second corpus.

Use detection-error trade-off curves and equal error rate to compare the selected systems on a different speech corpus.

03

Compare against GMM baselines.

Use diagonal-covariance Gaussian mixture models for vowels and speaker identification, comparing mixture complexity, utterance aggregation and UBM-MAP adaptation.

03 / RESULTS

Results.

23.77%No-context MLP · EER
15.73%GRU · EER
7.73%Tuned TCN · EER

Selected reported cross-corpus results over 495,486 test frames. The tuned TCN includes median post-processing. These figures are retained evaluation results, not a new benchmark run.

Context & limitations

Datasets, labels and checkpoints are excluded from the public repository. The GMM workflow needs a supplied or replacement MFCC frontend. Closed-set speaker results do not establish general-world identification performance.

04 / SUMMARY

The cross-corpus evaluation shows the effect of model family and temporal context; it is not a general speech-recognition claim.
Open repository
NEXT PROJECT / 06

Productiv / Genesys.

Team-built Rails productivity platform