Off-the-shelf speech recognition looks excellent on benchmarks. In your product, with your users, accents, microphones and vocabulary, it can be a different story. I help teams measure how well ASR really works for them and decide what to do about it.
What I offer
- Test set design: evaluation data that reflects your actual users, domains and recording conditions.
- Error analysis beyond WER: which errors matter for your task (names, numbers, intents) and which are harmless.
- Robustness: accents, dialects, non-native speech, children’s and older voices, noisy environments.
- Model selection and adaptation: comparing vendors and open models, and planning fine-tuning or custom vocabulary where it pays off.
- Data collection and annotation: transcription guidelines and quality control for training and test data.