We put your name into the data AI learns from.
Speech AI learns from patterns in its training data. A name with non-English origin or a unique spelling is rare in those patterns, so voice assistants mishear it, mispronounce it, and send the customer who asked for you somewhere else. EarlGrid fixes this at the source, so that AI hears your name and says your name.
How it runs
3 weeks of our work. Then the platforms take theirs.
1 Tell us the name. How you say it, and how you'll accept it being said. We record the baseline before we touch anything.
2 We build the records. Text, audio, and pronunciation in every format the platforms take: IPA, respellings, phoneme tags, and the data files AI learns from.
3 We submit them. To the open datasets and voice assistants on our list, with receipts.
4 We test it three times. Same recordings, same method: before the fix, at week 3, and at month 6. Results in writing.
What you get
1 Training data fix. Your name, in text and audio, published upstream in the open datasets speech AI learns from.* (Speech-to-text.)
2 Pronunciation data submission. A pronunciation record for every name, submitted to each voice assistant and lexicon that accepts one.** (Text-to-speech.)
3 An Alexa skill submission. An Alexa skill, with your name as its invocation name.
4 3 recognition reports. Same recordings and same method before the fix, at week 3, and at month 6.
* Mozilla Common Voice, Wikimedia Commons via Lingua Libre, and Wikidata. (+Wikipedia if the brand already qualifies)
** Wikidata (IPA), CMUdict and open lexicons, Alexa (via skill), Samsung Bixby, and Naver CLOVA. (+Siri if the brand has an iOS app)
