Improve transcription accuracy with hinting

1

Voci’s speech-to-text engine provides high levels of accuracy with no customization. From this strong baseline, it is possible to improve accuracy further by tuning the ASR to better understand each enterprise's specific call interactions with its customers. Voci Labs is developing a hinting solution to help drive the tuning process.

Tuning Options

hinting-diagram_Artboard 2No TuningHintingSubstitutionLight Language ModelingFull Custom Language ModelingFull Custom Language and Acoustic ModelingLess effortLess costLess tunedMore effortMore costMore tuned
Hinting-mobileNo TuningHintingSubstitutionLight Language ModelingFull Custom Language ModelingFull Custom Language and Acoustic ModelingLess effortLess costLess tunedMore effortMore costMore tuned

A “hint” is an important word or short phrase that you expect to be used frequently in a specific call or set of calls, such as the names of brands, products, companies, people, and phone numbers specific to the call environment. Because those words do not occur together with high frequency in general speech, Voci's off-the-shelf language models consider such word pairings to be unlikely. Without hinting, these phrases will be transcribed at lower accuracy than the rest of the conversation.

Hinting is an effective method for improving the accuracy of targeted phrases, such as those directly impacting categorization and other types of analytics. Hinting is also useful for correcting frequently occurring phrases with unusual word groupings, such as those found in marketing slogans, disclaimers, and required legal statements.

Voci Labs projects are still being developed. Interested in putting them into production? Get in touch.