Data Collection and Transcription

Speech recognition systems can benefit from application-specific training data collected from diverse subject pools, and from evaluation of recognizer performance through detailed transcription of user interactions and analysis of recognition errors. SRI International's Speech Technology and Research Labortary is a leader in providing these speech data collection and transcription services for both government and commercial applications.

Data Collection

SRI has facilities for collecting analog and digital telephone speech and stereo channel 16-bit workstation speech from our vast collection of low-to-high quality microphones. We have the ability to collect data from very large and diverse subject pools while maintaining specified demographic balances. We collect and transcribe native and non-native foreign speech in addition to native and non-native English speech. We also design interaction materials for data collection to include such prompt styles as phonetically balanced read speech, spontaneous responses, and conversational speech.

Transcription

SRI provides verbatim transcriptions of speech data which includes markings for disfluencies, non-linguistic events and mispronunciations. All data is checked for accuracy and accuracy rates are guaranteed. All transcription is performed by professional transcribers native to the language of speech being transcribed.

Evaluation

SRI provides analysis of recognition system performance through detailed breakdown of recognition errors. Misrecognitions can be categorized according to speaker, gender, dialect and background noise conditions.