Skip to content
heiSay hello

Vietnamese ASR: Parameter-efficient Speech Recognition

A CNN and residual BiLSTM model for Vietnamese speech recognition with CTC decoding.

In progress

Co-researcher / co-author

Sound becomes language.CNNBiLSTMAa5.4M parameters745-hour Vietnamese speech corpus
Method schematic · Not measured model output
Parameters
5.4M
Speech corpus
745 h
Reported WER
33.09%

Problem

Recognising Vietnamese speech with a compact end-to-end model. This research explores a CNN frontend and a residual recurrent encoder rather than assuming that a larger model is always the right starting point.

Method

  • Four CNN layers process the input speech features.
  • Three residual BiLSTM layers model the temporal sequence.
  • CTC connects variable-length speech inputs with text outputs.
  • The documented model has 5.4 million parameters and uses a 745-hour Vietnamese speech corpus.

Contribution

I work on this project as a co-researcher and co-author. The portfolio documents the model architecture, corpus scale and reported recognition results.

Reported evaluation

The original project reports 33.09% word error rate (WER) and 15.28% character error rate (CER). Lower is better for both measures.

Limitations

The portfolio does not include a complete benchmark protocol, dataset split specification or comparison table. These numbers should be read as reported project results, not evidence of state-of-the-art performance. The work remains in progress.