A team fine-tunes an NVIDIA Riva ASR pipeline on domain-specific vocabulary but finds the word error rate remains high on rare technical terms even after acoustic model fine-tuning. Which component should they target next to reduce errors on those terms?
- AThe vocoder, because it synthesises output waveforms and controls pronunciation fidelity for technical terms.
- BThe mel spectrogram extractor, because adjusting the filterbank resolution improves how rare phonemes are represented in the feature space.
- CThe beam-search width parameter, because widening the search ensures more candidate tokens are explored per decoding step.
- DThe language model component, because it assigns probabilities to word sequences and can be updated with a custom vocabulary or n-gram model. Correct
Why A is wrong: Vocoders are a TTS component that converts mel spectrograms into audio waveforms; they have no role in ASR decoding or vocabulary coverage.
Why B is wrong: Mel spectrogram extraction is a fixed pre-processing step; changing filterbank parameters affects all phonemes equally and does not address vocabulary coverage gaps.
Why C is wrong: Increasing beam width can marginally improve recall of low-probability paths, but it does not introduce domain vocabulary knowledge and has diminishing returns well before fixing a vocabulary-coverage deficit.
Why D is correct: In a hybrid ASR system the language model scores candidate transcriptions; injecting domain vocabulary and phrase probabilities directly reduces substitution errors on rare terms the acoustic model already produces as phoneme sequences.