A computationally efficient model for multipitch and periodicity analysis of complex audio signals is presented. The model essentially divides the signal into two channels, below and above 1000 Hz, computes a "generalized" autocorrelation of the low-channel signal and of the envelope of the high-channel signal, and sums the autocorrelation functions. The summary autocorrelation function (SACF) is further processed to obtain an enhanced SACF (ESACF). The SACF and ESACP representations are used in observing the periodicities of the signal. The model performance is demonstrated to be comparable to those of recent time-domain models that apply a multichannel analysis. In contrast to the multichannel models, the proposed pitch analysis model can be run in real time using typical personal computers. The parameters of the model are experimentally tuned for best multipitch discrimination with typical mixtures of complex tones. The proposed pitch analysis model may be used in complex audio signal processing applications, such as sound source separation, computational auditory scene analysis, and structural representation of audio signals. The performance of the model is demonstrated by pitch analysis examples using sound mixtures which are available for download at http://www.acoustics.hut.fi/-ttolonen/pitchAnalysis/.

A nonlinear discrete-time model that simulates a vibrating string exhibiting tension modulation nonlinearity is developed. The tension modulation phenomenon is caused by string elongation during transversal vibration. Fundamental frequency variation and coupling of harmonic modes are among the perceptually most important effects of this nonlinearity. The proposed model extends the linear bidirectional digital waveguide model of a string. It is also formulated as a computationally more efficient single-delay-loop structure. A method of reducing the computational load of the string elongation approximation is described, and a technique of obtaining the tension modulation parameter from recorded plucked string instrument tones is presented. The performance of the model is demonstrated with analysis/synthesis experiments and with examples of synthetic tones.

A model for multi-pitch and periodicity analysis of complex audio signals is presented that is more efficient and practical than the Meddis and O'Mard (see J. Acoust. Soc. Am., vol.102, p.1811-20, 1997) unitary pitch perception model, yet exhibits very similar behavior. We also demonstrate how to apply this model to source separation of complex audio signals such as polyphonic and multi-instrumental music and mixtures of simultaneous speakers. Such analysis techniques are important for automatic transcription of music and structural representation of audio signals.

Digital waveguide modeling of a nonlinear vibrating string is investigated when the nonlinearity is essentially caused by tension modulation. We derive synthesis models where the nonlinearity is implemented with a time-varying fractional delay filter. Also, conversion from a dual-delay-line physical model into a single-delay-loop model is explained. Realistic synthetic tones with nonlinear effects are obtained by introducing minor amendments to a linear string synthesis algorithm. It is shown how synthetic plucked-string tones are modified as a consequence of tension modulation.

A technique for rapid speaker adaptation, called eigenvoices, was introduced recently. The key idea is to confine models in a very low-dimensional linear vector space. This space summarizes a priori knowledge that we have about speaker models. In many practical systems, however, there is a mismatch between the conditions in which the training data were collected and test conditions: prior knowledge becomes improper. Furthermore, prior statistics or models of this mismatch may not be available. We expose two key results: first, we use a maximum-likelihood estimator of prior information in matched conditions, called MLES, leading to an improvement of adaptation by a relative 14%, and second, we show how we can apply a blind scheme for learning noise, MLLR, achieving an additional 7:7% relative improvement in noisy conditions.