ETSI ES 202 050-2007 Speech Processing Transmission and Quality Aspects (STQ) Distributed speech recognition Advanced front-end feature extraction algorithm Compression algorithms .pdf

1、 ETSI ES 202 050 V1.1.5 (2007-01)ETSI Standard Speech Processing, Transmission and Quality Aspects (STQ);Distributed speech recognition;Advanced front-end feature extraction algorithm;Compression algorithmsfloppy3 ETSI ETSI ES 202 050 V1.1.5 (2007-01) 2 Reference RES/STQ-00108a Keywords algorithm, s

2、peech ETSI 650 Route des Lucioles F-06921 Sophia Antipolis Cedex - FRANCE Tel.: +33 4 92 94 42 00 Fax: +33 4 93 65 47 16 Siret N 348 623 562 00017 - NAF 742 C Association but non lucratif enregistre la Sous-Prfecture de Grasse (06) N 7803/88 Important notice Individual copies of the present document

3、 can be downloaded from: http:/www.etsi.org The present document may be made available in more than one electronic version or in print. In any case of existing or perceived difference in contents between such versions, the reference version is the Portable Document Format (PDF). In case of dispute,

4、the reference shall be the printing on ETSI printers of the PDF version kept on a specific network drive within ETSI Secretariat. Users of the present document should be aware that the document may be subject to revision or change of status. Information on the current status of this and other ETSI d

5、ocuments is available at http:/portal.etsi.org/tb/status/status.asp If you find errors in the present document, please send your comment to one of the following services: http:/portal.etsi.org/chaircor/ETSI_support.asp Copyright Notification No part may be reproduced except as authorized by written

6、permission. The copyright and the foregoing restriction extend to reproduction in all media. European Telecommunications Standards Institute 2007. All rights reserved. DECTTM, PLUGTESTSTM and UMTSTM are Trade Marks of ETSI registered for the benefit of its Members. TIPHONTMand the TIPHON logo are Tr

7、ade Marks currently being registered by ETSI for the benefit of its Members. 3GPPTM is a Trade Mark of ETSI registered for the benefit of its Members and of the 3GPP Organizational Partners. ETSI ETSI ES 202 050 V1.1.5 (2007-01) 3 Contents Intellectual Property Rights5 Foreword.5 Introduction 5 1 Sc

8、ope 6 2 References 6 3 Definitions, symbols and abbreviations .7 3.1 Definitions7 3.2 Symbols8 3.3 Abbreviations .8 4 System overview 9 5 Feature Extraction Description.10 5.1 Noise Reduction .10 5.1.1 Two stage mel-warped Wiener filter approach.10 5.1.2 Buffering.11 5.1.3 Spectrum estimation .11 5.

9、1.4 Power spectral density mean.12 5.1.5 Wiener filter design 13 5.1.6 VAD for noise estimation (VADNest)14 5.1.7 Mel filter-bank16 5.1.8 Gain factorization .17 5.1.9 Mel IDCT .18 5.1.10 Apply filter19 5.1.11 Offset compensation .20 5.2 Waveform Processing.20 5.3 Cepstrum Calculation.21 5.3.1 Log en

10、ergy calculation21 5.3.2 Pre-emphasis (PE) 21 5.3.3 Windowing (W)22 5.3.4 Fourier transform (FFT) and power spectrum estimation.22 5.3.5 Mel filtering (MEL-FB)22 5.3.6 Non-linear transformation (Log).24 5.3.7 Cepstral coefficients (DCT)24 5.3.8 Cepstrum calculation output .24 5.4 Blind Equalization2

11、4 5.5 Extension to 11 kHz and 16 kHz sampling frequencies .25 5.5.1 FFT-based spectrum estimation25 5.5.2 Mel filter-bank26 5.5.3 High-frequency band coding and decoding 27 5.5.4 VAD for noise estimation and spectral subtraction in high-frequency bands.28 5.5.5 Merging spectral subtraction bands wit

12、h decoded bands29 5.5.6 Log energy calculation for 16 kHz .30 6 Feature Compression30 6.1 Introduction 30 6.2 Compression algorithm description30 6.2.1 Input30 6.2.2 Vector quantization.31 7 Framing, Bit-Stream Formatting and Error Protection.32 7.1 Introduction 32 7.2 Algorithm description.32 7.2.1

13、 Multiframe format 32 7.2.2 Synchronization sequence.33 ETSI ETSI ES 202 050 V1.1.5 (2007-01) 4 7.2.3 Header field 33 7.2.4 Frame packet stream .34 8 Bit-Stream Decoding and Error Mitigation35 8.1 Introduction 35 8.2 Algorithm description.35 8.2.1 Synchronization sequence detection .35 8.2.2 Header

14、decoding .35 8.2.3 Feature decompression .35 8.2.4 Error mitigation 36 8.2.4.1 Detection of frames received with errors 36 8.2.4.2 Substitution of parameter values for frames received with errors.36 9 Server Feature Processing 39 9.1 lnE and c(0) combination .39 9.2 Derivatives calculation.39 9.3 Fe

15、ature vector selection39 Annex A (informative): Voice Activity Detection 40 A.1 Introduction 40 A.2 Stage 1 - Detection .40 A.3 Stage 2 - VAD Logic42 Annex B (informative): Bibliography.44 History 45 ETSI ETSI ES 202 050 V1.1.5 (2007-01) 5 Intellectual Property Rights IPRs essential or potentially e

16、ssential to the present document may have been declared to ETSI. The information pertaining to these essential IPRs, if any, is publicly available for ETSI members and non-members, and can be found in ETSI SR 000 314: “Intellectual Property Rights (IPRs); Essential, or potentially Essential, IPRs no

17、tified to ETSI in respect of ETSI standards“, which is available from the ETSI Secretariat. Latest updates are available on the ETSI Web server (http:/webapp.etsi.org/IPR/home.asp). Pursuant to the ETSI IPR Policy, no investigation, including IPR searches, has been carried out by ETSI. No guarantee

18、can be given as to the existence of other IPRs not referenced in ETSI SR 000 314 (or the updates on the ETSI Web server) which are, or may be, or may become, essential to the present document. Foreword This ETSI Standard (ES) has been produced by ETSI Technical Committee Speech Processing, Transmiss

19、ion and Quality Aspects (STQ), and is now submitted for the ETSI standards Membership Approval Procedure. Introduction The performance of speech recognition systems receiving speech that has been transmitted over mobile channels can be significantly degraded when compared to using an unmodified sign

20、al. The degradations are as a result of both the low bit rate speech coding and channel transmission errors. A Distributed Speech Recognition (DSR) system overcomes these problems by eliminating the speech channel and instead using an error protected data channel to send a parameterized representati

21、on of the speech, which is suitable for recognition. The processing is distributed between the terminal and the network. The terminal performs the feature parameter extraction, or the front-end of the speech recognition system. These features are transmitted over a data channel to a remote “back-end

22、“ recognizer. The end result is that the degradation in performance due to transcoding on the voice channel is removed and channel invariability is achieved. The present document presents a standard for a front-end to ensure compatibility between the terminal and the remote recognizer. The first ETS

23、I standard DSR front-end ES 201 108 1 was published in February 2000 and is based on the Mel-Cepstrum representation that has been used extensively in speech recognition systems. This second standard is for an Advanced DSR front-end that provides substantially improved recognition performance in bac

24、kground noise. Evaluation of the performance during the selection of this standard showed an average of 53 % reduction in speech recognition error rates in noise compared to ES 201 108 1. ETSI ETSI ES 202 050 V1.1.5 (2007-01) 6 1 Scope The present document specifies algorithms for advanced front-end

25、 feature extraction and their transmission which form part of a system for distributed speech recognition. The specification covers the following components: - the algorithm for advanced front-end feature extraction to create Mel-Cepstrum parameters; - the algorithm to compress these features to pro

26、vide a lower data transmission rate; - the formatting of these features with error protection into a bitstream for transmission; - the decoding of the bitstream to generate the advanced front-end features at a receiver together with the associated algorithms for channel error mitigation. The present

27、 document does not cover the “back-end“ speech recognition algorithms that make use of the received DSR advanced front-end features. The algorithms are defined in a mathematical form or as flow diagrams. Software implementing these algorithms written in the C programming language is contained in the

28、 ZIP file es_202050v010105p0.zip which accompanies the present document. Conformance tests are not specified as part of the standard. The recognition performance of proprietary implementations of the standard can be compared with those obtained using the reference C code on appropriate speech databa

29、ses. It is anticipated that the DSR bitstream will be used as a payload in other higher level protocols when deployed in specific systems supporting DSR applications. In particular, for packet data transmission, it is anticipated that the IETF AVT RTP DSR payload definition (see bibliography) will b

30、e used to transport DSR features using the frame pair format described in clause 7. The Advanced DSR standard is designed for use with discontinuous transmission and to support the transmission of Voice Activity information. Annex A describes a VAD algorithm that is recommended for use in conjunctio

31、n with the Advanced DSR standard, however it is not part of the present document and manufacturers may choose to use an alternative VAD algorithm. 2 References The following documents contain provisions which, through reference in this text, constitute provisions of the present document. References

32、are either specific (identified by date of publication and/or edition number or version number) or non-specific. For a specific reference, subsequent revisions do not apply. For a non-specific reference, the latest version applies. Referenced documents which are not found to be publicly available in

33、 the expected location might be found at http:/docbox.etsi.org/Reference. NOTE: While any hyperlinks included in this clause were valid at the time of publication ETSI cannot guarantee their long term validity. 1 ETSI ES 201 108: “Speech Processing, Transmission and Quality aspects (STQ); Distribute

34、d speech recognition; Front-end feature extraction algorithm; Compression algorithms“. 2 ETSI EN 300 903: “Digital cellular telecommunications system (Phase 2+) (GSM); Transmission planning aspects of the speech service in the GSM Public Land Mobile Network (PLMN) system (GSM 03.50)“. ETSI ETSI ES 2

35、02 050 V1.1.5 (2007-01) 7 3 Definitions, symbols and abbreviations 3.1 Definitions For the purposes of the present document, the following terms and definitions apply: analog-to-digital conversion: electronic process in which a continuously variable (analog) signal is changed, without altering its e

36、ssential content, into a multi-level (digital) signal blind equalization: process of compensating the filtering effect that occurs in signal recording NOTE: In the present document blind equalization is performed in the cepstral domain. DC-offset: direct current (DC) component of the waveform signal

37、 discrete cosine transform: process of transforming the log filter-bank amplitudes into cepstral coefficients fast fourier transform: fast algorithm for performing the discrete Fourier transform to compute the spectrum representation of a time-domain signal feature compression: process of reducing t

38、he amount of data to represent the speech features calculated in feature extraction feature extraction: process of calculating a compact parametric representation of speech signal features which are relevant for speech recognition NOTE: The feature extraction process is carried out by the front-end

39、algorithm. feature vector: set of feature parameters (coefficients) calculated by the front-end algorithm over a segment of speech waveform framing: process of splitting the continuous stream of signal samples into segments of constant length to facilitate blockwise processing of the signal frame pa

40、ir packet: definition is specific to ES 202 050: the combined data from two quantized feature vectors together with 4 bits of CRC front-end: part of a speech recognition system which performs the process of feature extraction magnitude spectrum: absolute-valued Fourier transform representation of th

41、e input signal multiframe: grouping of multiple frame vectors into a larger data structure mel-frequency warping: process of non-linearly modifying the frequency scale of the Fourier transform representation of the spectrum mel-frequency cepstral coefficients: cepstral coefficients calculated from t

42、he mel-frequency warped Fourier transform representation of the log magnitude spectrum notch filtering: filtering process in which the otherwise flat frequency response of the filter has a sharp notch at a predefined frequency NOTE: In the present document, the notch is placed at the zero frequency,

43、 to remove the DC component of the signal. offset compensation: process of removing DC offset from a signal power spectral density: squared magnitude spectrum of the signal pre-emphasis: filtering process in which the frequency response of the filter has emphasis at a given frequency range NOTE: In

44、the present document, the high-frequency range of the signal spectrum is pre-emphasized. sampling rate: number of samples of an analog signal that are taken per second to represent it digitally ETSI ETSI ES 202 050 V1.1.5 (2007-01) 8 SNR-dependent Waveform Processing (SWP): processing of signal wave

45、form with objective to emphasize high-SNR waveform portions and de-emphasize low-SNR waveform portions voice activity detection: process of detecting voice activity in the signal NOTE: In the present document one voice activity detector is used for noise estimation and a second one is used for non-s

46、peech frame dropping. wiener filtering: filtering of signal by using Wiener filter (filter designed by using Wiener theory). NOTE: In this work, objective of Wiener filtering is to de-noise signal. windowing: process of multiplying a waveform signal segment by a time window of given shape, to emphas

47、ize pre-defined characteristics of the signal zero-padding: method of appending zero-valued samples to the end of a segment of speech samples for performing a FFT operation 3.2 Symbols For the purposes of the present document, the following symbols apply: For feature extraction: bin FFT frequency in

48、dex c(i) cepstral coefficients; used with appropriate subscript E(k) filter-bank energy; used with appropriate subscript H(bin) or H(k) Wiener filter frequency characteristic; used with appropriate subscript h(n) Wiener filter impulse response; used with appropriate subscript k filter-bank band inde

49、x KFBnumber of bands in filter-bank lnE log-compressed energy feature appended to cepstral coefficients n waveform signal time index N length, (e.g. frame length, FFT length, .); used with appropriate subscript P(bin) power spectrum; used with appropriate subscript S(k) log filter-bank energy; used with appropriate subscript s(n) waveform signal; used with appropriate subscript t frame time index TPSDnumber of frames used in the PSD Mean technique w(n) windowing function in time domain; used with appropriate subscript W(bin) frequency window X(bin) FFT complex