Modern neural TTS can sound remarkably natural, until it reads a date wrong, mispronounces a product name or picks the wrong reading of a word like “lead” or “record”. These errors are what users notice and remember. I help teams find and fix them.
What I offer
- Linguistic preprocessing: text normalization (numbers, dates, currencies, abbreviations), homograph disambiguation, custom pronunciation dictionaries, grapheme-to-phoneme models and language model fine-tuning for these tasks.
- Multilingual and cross-lingual voices: handling code-switching, foreign names and loanwords, and dialectal variation.
- Prosody and expressive speech: conversational, expressive and long-form reading styles.
- Speech database design and recording: script design balanced for sentence types, speaking styles, intonation and cross-lingual phenomena, collected with informed consent and fair compensation.
- Evaluation: listening tests, automatic MOS prediction and ASR-based intelligibility checks, designed around the errors that matter for your application.
Background
I spent many years building production TTS as a Senior Research Scientist at ReadSpeaker, after working as an Assistant Professor at the Center for Spoken Language Understanding (CSLU) at OHSU. For more context, read What is text-to-speech synthesis? and my blog.