Integration of the harmonic plus noise model (HNM) into the hidden Markov model-based speech synthesis system (Master Thesis)
(2006)
Files
MasterThesis_Hemptinne.pdf
Open Access - Adobe PDF
- 1.38 MB
Details
- Authors
- Abstract
- In the present project, we developed and tested a new model-based text-to-speech (TTS) system, integrating the Harmonic plus Noise Model (HNM) into the Hidden Markov Model-based Speech Synthesis System (HTS). This integration leads to a TTS system that requires smaller development time and cost, in comparison with the usual state-of-the-art TTS systems typically based on automatic selection and synthesis of subword units (e.g., diphones), while also producing a better-quality speech output (compared to HTS alone). This quality enhancement is achieved by replacing the source filter modeling approach typically used in HTS with the HNM model, which is known for being able to synthesize natural sounding speech under various prosodic modifications. The basic idea behind HNM is to model speech as being composed of harmonic and noise parts. Voiced frames comprise a harmony part and a noise part, separated by the time-varying maximum voiced frequency, whereas unvoiced frames are only composed of a noise part. The HNM algorithm consists in two steps: (1) the HNM analysis, i.e., the computation of the HNM parameters of every acoustic unit of the training database, and (2) the HNM synthesis, i.e., the speech waveform synthesis from the HNM parameters. The HTS system comprises a training and a synthesis part. The training part consists in com-puting the parameters modeling the database and in training context-dependent HMMs. During the synthesis part, given the target prosodic and phonetic labels corresponding to the text to synthesize, the adequate context dependent HMMs are concatenated to build a composite HMM. The most likely parameters are then estimated and used to synthesize speech waveform using a filter-based approach. In the TTS system develop ed in this work, the utterance database is modeled by HNM parameters, which constitutes the first modification of the HTS system. These parameters and their dynamic features are used to train context dependent HMMs, like in HTS. A composite HMM is then constructed and the HMM parameters are generated by maximum likelihood estimation. Finally, the speech waveform is obtained by HNM synthesis, which is the second modification of the HTS system. This constitutes the general description of the TTS system used in this work, which has been implemented in three different ways: (1) The HNM parameters are extra ted at a fixed rate from the training database; they in lude the linear predictive cepstral coefficients (LPCC) and the fundamental frequency; (2) The same parameters are extracted pitch-synchronously; (3) The extraction is again pitch-synchronous, but maximum voiced frequency is also modeled. In conclusion, a slightly better speech waveform quality is obtained in the third case.
- Affiliations
Citations
Hemptinne, C. (2006). Integration of the harmonic plus noise model (HNM) into the hidden Markov model-based speech synthesis system (Master Thesis). https://hdl.handle.net/2078.5/28145