Reproduce “A comparative analysis between Conformer-Transducer, Whisper, and wav2vec2 for improving the child speech recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Andrei Barcovschi, Rishabh Jain, Peter Corcoran
The C3 Imaging Research Center at the University of Galway conducted a comparative analysis of Conformer-transducer, Whisper, and wav2vec2 for child speech recognition. Their study found that finetuned wav2vec2-large models consistently achieved the lowest Word Error Rates on child speech datasets, reporting 7.42% on MyST_test and 2.99% on PFS_test.
Key insights
- Domain-specific finetuning is critical for significantly improving ASR performance on child speech, substantially reducing WERs across all evaluated models.
- Wav2vec2's self-supervised pretraining provides a robust foundation that generalizes exceptionally well to child speech, achieving superior performance with comparatively less finetuning data.
- Conformer-transducer models exhibit strong baseline performance at smaller scales but struggle with generalization and robustness to noise when scaled up or finetuned on limited, single-domain child speech datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.