Wideband Speech Recovery from Narrowband Speech Using Classified Codebook Mapping

Speech sounds occupy 8 kHz or more of bandwidth. However, current public telephone networks limit the speech bandwidth to 300–3400 Hz. Telephone speech is characterized by thin and muffled sounds, and degraded speaker identification. We describe an algorithm which generates the missing highband components from the narrowband speech signal. The algorithm is based on three acoustic-phonetic classified narrowband-to-wideband linear prediction (LP) spectrum mapping codebooks to recover the missing highband spectrum. Subjective tests show that the reconstructed wideband speech improves the speech quality. LP spectrum bandwidth expansion is used to avoid sharp spectral peaks. The mean SD (log-spectrum distortion) decreases by 0.93 dB, comparing to non-classified codebooks without LP bandwidth expansion. INTRODUCTION Current public telephone networks provide voice services for narrowband speech (300–3400 Hz). Since the human voice occupies 8 kHz or more in bandwidth, telephone speech is characterized by thin and muffled sounds. That results in a degradation of the intelligibility of speech. For example, it is very difficult to distinguish between phonemes, /s/ and /f/, because their highband components (over 3.4 kHz) are important discriminators. Most other phonemes have lost, to some extent, their fidelity in telephony speech. In the coming 3G wireless communications systems with the advanced speech coding technology, Adaptive Multi-Rate codec (AMR) will be employed to provide wideband speech coding. However, the connection between the existing PSTN and 3G wireless systems deteriorates the quality of the service because of the bandwidth bottleneck of PSTN. A promising solution is to use a wideband recovery technique to improve the service quality of PSTN when the signal passes to a wideband network. Several attempts have been made to address the wideband recovery issue. These methods generate an excitation signal which passes through a synthesis filter. There are two main issues for the reconstruction of the missing highband components: the reconstruction of the highband excitation and the highband spectral envelope (via the synthesis filter). One approach is to use a mapping between a codebook of narrowband spectra and a codebook of wideband spectra [Epps, 1998], [Jax, 2002]. However, since different voice sounds demonstrate a large divergence of their spectrum envelope characteristics between the narrowband and wideband speech, a single mapping codebook can result in large spectrum mapping errors. Our new approach for highband spectrum recovery is based on three acoustic-phonetic classified narrowband to highband spectrum mapping codebooks. The missing highband excitation can be regenerated by a bandpass envelope modulated Gaussian white noise with a mapping modulation gain [Qian & Kabal, 2002]. We first introduce the acoustic-phonetic classification of speech sounds. Then, we describe the wideband recovery system with three classified lowband to highband spectral envelope mapping codebooks and three excitation gain mapping codebooks. The wideband reconstruction simulation results of objective measurements for mean spectrum distortion (SD) and spectrograms are given in the last section. ACOUSTIC PHONETICS FOR CLASSIFICATION Acoustic-phonetics describes distinctive waveform, spectrum and power properties of speech sounds, or phonemes. We have considered two parameters, the high band (4 kHz to 8 kHz) vs. lowband (below 4 kHz) energy ratio, r g and the voicing periodicity, or the pitch predictor gain β , to classify phonemes. The r g indicates, the feature of the energy of the missing highband. The parameter β is a measure of the degree of the waveform periodicity or the harmonic structure. It also correlates with Qian et al. Wideband Speech Recovery Proceedings of the 9th Australian International Conference on Speech Science & Technology Melbourne, December 2 to 5, 2002.  Australian Speech Science & Technology Association Inc. Accepted after abstract review page 107 the highband spectrum envelope. For our application, the 42 phonemes of American English can be classified into 3 groups using the two parameters. • Five unvoiced fricative phonemes: /s/, /f/, /θ /, /sh/, / ∫ /, the whisper, /h/ and 2 affricatives, /j/, /t ∫ /. They have large highband to lowband energy ratios and no harmonics in spectrum (small pitch gain). The parameter r g varies between 9.0 and 25.0 dB. The parameter β is in the range 0.1–0.25. A typical spectral envelope of the phoneme /s/ is shown in Fig. 1. • Four voiced fricatives and 6 stop explosive consonant phonemes: /v/, /th/, /z/, /zh/ and /p/, /t/, /k/, /b/, /d/, /g/. Their highband to low-band energy ratio is moderate, in the range of -13.0–9.0 dB. They have, to certain degree, periodicity in the waveform. The parameter β is in the range 0.25–0.90. The voiced phonemes have β close to the upper boundary. The spectrum envelope of the phoneme /k/ is depicted in Fig. 2. • The other 11 vowels, 6 diphthongs, 4 semivowels and 3 nasal consonant phonemes manifest a small lowband to highband energy ratio ( r g < –15 dB) and very good periodicity in the waveform (pitch gain 0.9 β > ). The phoneme /o/ spectrum envelope is illustrated in Fig. 3. The voicing degree parameter β plays an important role in the classification. We have used the voicing periodicity parameter to approximately classify the speech frames into three groups: 0.25 unvoiced phonemes 0.25 0.9 mixed phonemes 0.9 voiced phonemes β β β ≤ < < ≤ HIGHBAND RESTORATION The key part of a wideband reconstruction system is the highband spectrum envelope reproduction, as depicted in Fig. 4.