Frequency domain coding of speech

Tribolet, José; Crochiere, R.

doi:10.1109/tassp.1979.1163283

Cited by 132 publications

(34 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Denoting as X DF T [k] the DFT of a length 2N zeropadded version of x[n], it has been shown [4] that the envelope of the DCT is bounded by the envelope of the zeropadded DFT and in fact they are exactly related by…”

Section: Discrete Cosine Transform (Dct)mentioning

confidence: 99%

Frequency-domain linear prediction for temporal features

Athineos

Ellis

2003 IEEE Workshop on Automatic Speech Recognition and Understanding (IEEE Cat. No.03EX721)

View full text Add to dashboard Cite

Current speech recognition systems uniformly employ short-time spectral analysis, usually over windows of 10-30 ms, as the basis for their acoustic representations. Any detail below this timescale is lost, and even temporal structure above this level is usually only weakly represented in the form of deltas etc. We address this limitation by proposing a novel representation of the temporal envelope in different frequency bands by exploring the dual of conventional linear prediction (LPC) when applied in the transform domain. With this technique of frequency-domain linear prediction (FDLP), the 'poles' of the model describe temporal, rather than spectral, peaks. By using analysis windows on the order of hundreds of milliseconds, the procedure automatically decides how to distribute poles to best model the temporal structure within the window. While this approach offers many possibilities for novel speech features, we experiment with one particular form, an index describing the 'sharpness' of individual poles within a window, and show a large relative word error rate improvement from 4.97% to 3.81% in a recognizer trained on general conversational telephone speech and tested on a small-vocabulary spontaneous numbers task. We analyze this improvement in terms of the confusion matrices and suggest how the newlymodeled fine temporal structure may be helping.

show abstract

Section: Discrete Cosine Transform (Dct)mentioning

confidence: 99%

Frequency-domain linear prediction for temporal features

Athineos

Ellis

2003 IEEE Workshop on Automatic Speech Recognition and Understanding (IEEE Cat. No.03EX721)

View full text Add to dashboard Cite

show abstract

“…It has been reported [15] that good communication quality speech can be produced at 7.2 kbit/s using a vocoder-driven adaptive transform coding technique [6,84] and that similar quality speech can be obtained at 9.6 kbit/s using sub-band coding. Based on a relative count of the number of logic gates used, the complexity of the adaptive transform coder is estimated [15] to be approximately ten times that of the sub-band coder implemented using charge-couple-device filters.…”

Section: Motivation Fob Hybrid Frequency-domain Codingmentioning

confidence: 99%

“…These lobes are often referred to [6] as 'spectral fine structure'. The spectral fine structure is dependent mainly on the-pitch period and the degree of voiced periodicity.…”

Section: Spectrummentioning

confidence: 99%

See 1 more Smart Citation

Hybrid frequency-domain coding of speech signals

Foo

Turner

1983

IEE Proc. F Commun. Radar Signal Process. UK

View full text Add to dashboard Cite

A new method of coding speech signals is introduced and its application to the low-bit-rate transmission of speech is examined. The new method, which is referred to as hybrid frequency-domain coding (HFDC), is based on a combination of sub-band coding and transform coding techniques and thus involves block coding.In one version of the HFDC scheme, a fixed data rate is used. In this version, blocks of speech signal are coded using a fixed number of bits. The technique is of relatively low complexity and results of subjective tests indicate that the quality of speech produced by the technique at 7.2 kbit/s is comparable with that produced by 19 kbit/s continuously variable slope delta modulation (CVSD); and that at 9.6 kbit/s, it is comparable with 23 kbit/s CVSD.In a further application of the HFDC scheme, a variable bit rate technique is used which takes account of the nonstationary nature of speech. Block coding is again used but. the number of bits allocated is made to depend on the characteristics of the block. Subjective tests have shown that. improved speech quality can be obtained at the expense of slightly increased end-to-end delay.Quadrature mirror filters are important in sub-band coding and a new method of designing these filters is described. The effects of finite-word-length on the performance of systems using quadrature mirror filters are considered and formulae are derived relating to the bounds and variances of the errors due to analog-to-digital conversion; quantization of the filter coefficients; roundoff in arithmetic computations and the quantization associated with compression encoding. The results obtained from extensive simulations are shown to be in excellent agreement with the theoretically derived values.Various other issues relating to the design of the HFDC scheme are considered. These include the analysis of errors resulting from the use of the block-floating-point fast Walsh-Hadamard transform, and the automatic classification of speech into voiced, unvoiced and silent intervals. Last but not least, I would like to express my thanks to Dr.Tay Eng Soon for his support and encouragement; and to the Ministry of Defence for granting the DSO scholarship and the paid study leave. 5

show abstract

“…Pioneer work on transform coding was done in the late 1970s by Zelinsky and Noll [29,30], Tribolet and Crochiere [31]. Although that work was mainly intended for coding of speech in the frequency domain, it is the basis of almost all state-of-the-art audio coders.…”

Section: Frequency Domain Codersmentioning

confidence: 99%

Perceptual coding of narrowband audio signals at 8 kbit/s

Najafzadeh-Azghandi

Kabal²

1997 IEEE Workshop on Speech Coding for Telecommunications Proceedings. Back to Basics: Attacking Fundamental Problems in Speec

View full text Add to dashboard Cite

New applications such as Internet broadcast and communications, consumer multimedia products, digital AM broadcast and satellite networks are emerging. Those applications require moderate audio quality without annoying artifacts at bit rates below 16 kbit/s. Although speech coders provide high speech quality at bit rates around 8 kbit/s, they perform poorly when encoding audio signals. In this thesis, we present a novel transform coding paradigm based on the characteristics of the human hearing system. The proposed encoder, i.e., Narrowband Perceptual Audio Coder (NPAC), can accommodate a wide range of narrowband audio inputs without annoying artifacts at bit rates down to 8 kbit/s.NPAC employs a variety of algorithms to remove the perceptually irrelevant parts and statistical redundancies of the input signal. The new algorithms used in NPAC include a perceptual error measure in training the codebooks and selecting the best codewords, perceptually-based bit allocation algorithms and an adaptive predictive scheme to vector quantize the scale factors.The proposed encoder has moderate complexity and delivers good quality for narrowband audio inputs at around 1 bit/sample. Informal subjective tests have been conducted to compare the performance of NPAC with an 8 kbit/s commercially-available audio coder.The tests results show that NPAC performs better for both music and speech inputs. iii RésuméDes nouvelles technologies telles que la diffusion par Internet, la diffusion AM numérique, et les réseaux satellites deviennent de plus en plus populaires et constituent la base de plusieurs nouvelles applications et produits multimédias. La réussite de ces produits sur la marché dépend de la qualité des signaux audio et vidéo ainsi que de la largeur de bande utilisée. Pour le signal audio, il est désirable que le débit soit en bas de 16 kbit/s tout en offrant une qualité acceptable, c'est-à-dire sans de distorsion remarquable.Il està noter que certains codeurs de parole permettent de transmettre le signal de parole au débit de 8 kbit/s avec une très bonne qualité. Toutefois, puisque ces codeurs profitent de la structure particulière de la parole, ils ne peuvent pas offrir la même qualité audio pour d'autres signaux comme la musique.Dans cette thèse, nous présentons une philosophie d'encodage des signaux audio qui tient compte de la structure du système auditif. Le codeur proposé se nomme Codeur Audio Perceptuelà bandeÉtroite (CAPE). CAPE permet d'encoder plusieurs types de signal audioà bandeétroite au débit de 8 kbit/s sans de distorsion remarquable.Plusieurs nouveaux algorithmes sont utilisés dans CAPE afins d'éliminer la redondance statistique ainsi que la partie sans importance perceptuel du signal d'entrée. Parmi les nouveautés de CAPE, il y a une mesure d'erreur perceptuelle qui est utilisée lors de l'entraînement des tableaux de quantification, et pour la sélection du meilleur vecteur de ces tableaux lors de l'encodage. De plus, l'allocation des bits pour les gains du spectre dans différentes bandes de fréque...

show abstract

Frequency domain coding of speech

Cited by 132 publications

References 35 publications

Frequency-domain linear prediction for temporal features

Frequency-domain linear prediction for temporal features

Hybrid frequency-domain coding of speech signals

Perceptual coding of narrowband audio signals at 8 kbit/s

Contact Info

Product

Resources

About