Background

Several attempts continue to improve Text Input technologies for mobile devices; well-known methods include (a) Tripple-Tapping (b) T9/iTap/EziText Prediction (c) Palm-Graffiti (d) Mini-Qwerty keyboards (e) On-Screen soft keypads (f) Handwriting Recognition (g) Optical Character Recognition and (h) touch screen interface. Unfortunately, mimicking the full-size keyboard onto a mobile device seems a daunting task.

Obviously, Speech-to-Text seems like a natural alternative: if one could simply speak into their device and have the Text magically appear on the screen. Several attempts have been made to address the speech-to-text problem acknowledged to be “the holy grail of computing”. Unfortunately, the so-called "Dictation" approaches have had limited success because of (a) infinite language perplexity of the problem, (b) background and channel noises, (c) varied pronunciations, (d) unacceptable speaker-training methods, and (e) lack of intuitive error-correction.

Predictive Speech-to-Text™ by TravellingWave

TravellingWave takes an innovative approach that enhances Predictive Text input using Speech Recognition. Specifically, Predictive Speech-to-Text technology predicts words using the acoustics (speech) rendered by a user in addition to the letters inputted by the user; traditional predictive text input systems rely only on letters. The result is faster and simpler prediction of text. The technology builds on top of existing text input methods. Thus in situations wherein Speaking is undesirable, Predictive Speech-to-Text turns into Standard Text Prediction. The overall system uses the following modules:

• VoicePredict Multimodal Platform (patent-pending) handles very large vocabularies and incorporates information from one mode (for instance letters typed using a keyped) to enhance speech recognition of spoken words. The system "intelligently" combines information from multiple modes to result in a simple, fast input method.

• Frequency Localized Temporal Speech Processing (Proprietary) is based on the company's RAGs (Rao-Aronov-Garafutdinov Speech-Processing) algorithm which in turn is based on published research^1-3 on compact features modeling the travelling wave phenomena in the human cochlea. This module extracts modulation information from speech, as opposed to traditional power spectrum analysis. This enables VoicePredict system to function robustly in noisy environments.

• Spellation (patent-pending) technology is based on the philosophy of extracting redundant acoustic information. It requires users to speak-and-spell as they dictate. This constraint seems minimal especially in the absence of a full-size keyboard. This approach yields significant improvements in recognition accuracy, compared to standard dictation approaches.

• Acoustic and Language Modeling (proprietary): Unlike dictation of standard documents (like letters, memos, and medical/legal reports) text messages (sms, email, im) tend to be very short. For instance sms-text is restricted to a bandwidth of 140-160 characters per message. Our techniques may be viewed as an optimization of standard modeling techniques, for the messaging application.