Distributed Speech Recognition Where is 358 Madison Avenue

Transcription

1 Distributed Speech Recognition Where is 358 Madison Avenue David Pearce Motorola Labs

2 Voice & Multimodal Multimodal-enabled Voice-enabled Services User enters commands via: SPEECH KEYPAD Screen OUT Audio OUT System responds: GRAPHIC TEXT S SPEECH SOUNDS Keypad IN Speech IN 2

3 Distributed Speech Recognition Content Servers Client Devices Conventional [Wireless] Packet Data Network IP Netwo rk Voice Gateway / Server: VoiceXML / mm Browser Speech Resources (ASR, TTS, etc.) Speech Coder Circuit Switched Mobile Voice Channel Speech Decoder ISDN ASR Front-end ASR Decoder ASR Front-end Packet Data Channel e.g. GPRS or CDMA 1x ASR Decoder 3

4 Benefits of Word Accuracy (%) Baseline error free strong medium weak GSM signal strength Improves performance over wireless channels Minimises impact of codec & channel errors Consistent performance over coverage area Improved performance in background noise 53% reduction in error rate EFR Coded Speech Ease of integration of combined speech and data applications Use packet data channel for both and other data 4

5 Standards Distributed Speech Recognition Advanced front-end (Oct 2002) Extended Advanced Front-end (Nov 2003) Speech Enabled Services Fixed point standard created selected as the recommended codec for SES (Approved June 04) IETF 3GPP2 RTP payload formats for Specifications standardised rfc4060 Speech Enabled Services New Work Item (Approved Jan 2005) 5

6 Advanced Front-end (ES ) Noise Robust Front-end Half error rate cf mel-cepstrum in background noise Double Wiener filtering noise suppression Waveform processing Blind equalisation Representation: 12 cepstral coeffs, C0, loge Compression gives bit rate of 4.8kbit/s Feature Extraction 8 & 16 khz VAD input signal Noise Reduction Waveform Processing Cepstrum Calculation Blind Equalization to feature compression 6

7 Extension (ES ) Enables Speech waveform reconstruction at server for human listening Adds 800bps containing pitch (total 5.6kbps): Assists recogniser with tonal language recognition (e.g. Mandarin, Cantonese) Speech In ETSI Standard Front-End MFCC & 4800 bps Back-End Pitch & Class Estimation Pitch & 800 bps C H A N N E L Pitch Tracking and Smoothing Tonal Information Speech Reconstruction Speech Out 7

8 Results of ASR vendor evaluations in 3GPP 8 khz Number of db tested AMR4.75 Average Absolute Performance Average Absolute Performance Average Improvement Digits % Sub-word % Tone confusability % Channel errors % Weighted Average 36% Extensive testing on 21 different speech databases Covering different languages, tasks and environments Tests performed with IBM and Scansoft commercial recognisers Results above are for low data-rate comparison for packet data (< 8kbit/s) 8

9 Packet Switched Channel Errors Robustness to block errors narrow-band (8kHz) Word accuracy (%) AMR 12.2 AMR Block error rate (%) Aurora-3 Italian speech database GPRS network simulation for distribution of errors 3GPP Feb

10 Coded speech vs (Aurora-3 Italian) AMR 4.75 Degradation Well matched % Med mismatch % High mismatch % Average % EVRC Degradation Well matched % Med mismatch % High mismatch % Average % 10

11 Distributed Multimodal Architecture Handset MM Gateway Content Server J2ME Application Application Front End Front End RTP/SIP & SIP GPRS or 3G Network RTP/SIP RTP & SIP Multimodal Multi-Modal VoiceXML Browser Browser ASR Decoder HTTP Multimodal Applications and content Handset device Input modalities (i.e.,, keypad input, pen entry) Output media (e.g., Visual rendering, Decoded speech output) Application Environment (Java or WAP Browser) Protocols (SIP / RTP, Multimodal remote control) Multimodal Gateway Decoder Multimodal VoiceXML browser Protocols Applications and content Content authoring Content delivery 11