Project 01 / 05 · Speech recognition
Implemented and testedSpeech-to-Text Robustness & ASR Evaluation
I built a transcript-evaluation pipeline that aligns hypotheses to reference speech, separates substitution/deletion/insertion errors and reports corpus-weighted WER across noise, telephony and code-switching slices.
A single WER number can hide where recognition breaks. A caller reading an account number through a narrowband phone connection has a different failure profile from a quiet-room dictation sample.
What I implemented
- Implemented Unicode-normalized word and character error rates with dynamic-programming edit alignment.
- Built slice-aware corpus aggregation with separate error counts for clean, noisy, telephony and code-switching examples.
- Built a JSONL evaluation CLI and deterministic test fixtures, so an actual recognizer adapter can be compared without changing the scoring contract.
Data and execution path
- 01
Reference transcript + model hypothesis
- 02
Unicode-aware normalization and tokenization
- 03
Dynamic-programming sequence alignment
- 04
Substitution / insertion / deletion counts
- 05
Corpus-weighted WER / CER and condition breakdown
- 06
Regression report keyed to model and dataset version
Experiment and results
The included four-pair illustrative corpus scores 4 word errors across 27 reference words (WER 14.8%). Those hypotheses are authored fixtures, not output from a trained recognizer. The scoring algorithm is implemented and unit tested; plugging in an ASR model is a separate integration.
These results are from the specified local dataset or offline rendering pipeline. I keep them separate from real speech-corpus and live-call metrics because the populations and failure modes differ.
System-level constraints
- Versioned decoder outputs and restricted transcripts are essential to comparing ASR runs without changing the evaluation target.
- WER and CER do not measure entity or task correctness. I treat an incorrect account digit as a separate operational failure.
- Robustness claims depend on representative channels, languages, code-switching, accents and human-reviewed ground truth, not illustrative text fixtures.
- Self-supervised or weakly supervised ASR results are only comparable on held-out, versioned evaluation data.
What the measurements establish
This implementation evaluates transcripts; it does not train, fine-tune or serve an ASR model. Actual accent, noise, code-switching and language performance require representative labeled recordings and speaker-separated test sets.
A model can understand the sentence and still get the account number wrong. The database tends to care about the number.