• What is RECOVC ?
• Applications
• Standardization
• Documents
• Speech Quality Demo
• IBM Publications
• Related web sites
• Contact Person

The Recognition Compatible Voice Coder (RECOVC) is a new IBM proprietary low bit rate speech compression algorithm. It is designed to work in applications where compressed speech is to be processed by an Automatic Speech Recognition (ASR) engine for converting the speech into text, or processed by a Speaker Recognition/Verification engine. Compressing speech using the low bit rate compression technologies available today causes a degradation in the recognition rates, especially for large vocabulary, continuous speech recognition tasks. RECOVC compresses speech while keeping the recognition rates intact. This is possible because of the development of a compression algorithm for the speech recognition feature vector, MFCC (Mel-Frequency Cepstral Coefficients), in which recognition rates are unimpaired and the development of a speech reconstruction method from the MFCC and pitch period.

The RECOVC technology can be used in Distributed Speech Recognition (DSR) systems as the preferred voice that carries format between a client that captures voice and a server that recognizes it. RECOVC technologies can be used in the following applications:

• Digital voice recorders and other Personal Digital Assistant (PDA) devices, which record voice that is later recognized.
• Accessing voice portals on the Web for information retrieval and other Interactive Voice Response (IVR) services from Internet phones, cellular phones, or other portable devices.
• Voice mail services, voice logging, and other speech database services with enhanced features such as word spotting, speaker identification, speech transcription, and other speech recognition-based techniques.
• Automatic transcription of voice messages for pagers.
• Speech recognition based services, where playback of the archived recordings is important for legal issues.

• Using RECOVC in a DSR environment 
• Using RECOVC technology in a small footprint concatenative TTS system

• ETSI STQ Aurora group develops standards for Distributed Speech Recognition (DSR) applications.
• RECOVC is a substantial part of the new working item: "Front-end extension for speech reconstruction and tonal language recognition". 

• RECOVC presentation slides (in PDF)

• Click here for an on-line demonstration of the speech reconstruction quality.
• Beware ! The files in the above demo are WAV files (16 - bit per sample, PCM) and therefore are quite big. If you prefer to download this demo once as a single package to your local machine, download the following ZIP file (~3.4 Mbyte), extract it, and play the files locally.

• D. Chazan, M. Zibulski, R. Hoory, and G. Cohen, "Efficient Periodicity Extraction Based on Sine-wave Representation and its Application to Pitch Determination of Speech Signals", in proc. of the 7th European Conference on Speech Communication and Technology (EUROSPEECH-2001),  Sept. 3-7, 2001, Aalborg Denmark.
• S. H. Maes, G. Cohen, R. Hoory, and D. Chazan, "Conversational Networking: Conversational Protocols for Transport, Coding and Control", in proc. 6th Int. Conf. Spoken Language Processing, Beijing China, Oct. 2000 (ICSLP-2000).
• D. Chazan, G. Cohen, R. Hoory, and M. Zibulski, "Low Bit Rate Speech Compression for Playback in Speech Recognition Systems", in proc. European Signal Processing Conference,  EUSIPCO 2000. Also presented in the Israeli IEEE Signal Processing Workshop in the Technion (Israel Institute of Technology), June 2000.
• D. Chazan, G. Cohen, R. Hoory, and M. Zibulski, "Speech Reconstruction from Mel-frequency Cepstral Coefficients and Pitch Frequency", in proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing, ICASSP 2000.
• G. N. Ramaswamy and P.S. Gopalakrishnan, "Compression of Acoustic Features for Speech Recognition in Network Environments", in proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing, ICASSP 1998.

• IBM Voice Systems 
• IBM Research, Haifa Lab.
• Voice Technology Initiative for Mobile Enterprise Solutions - VoiceTIMES 

Alex Sorin
IBM Haifa Research Lab.
Multimedia and Signal Processing Dept.
Tel: +972-4-8296-289
e-mail: sorin@il.ibm.com