Using noisy data for speaker adaptation in HMM-based Speech Synthesis

• Training a good-quality synthetic voice requires some hours of speech data recorded in good conditions.
• Using a speaker-adaptive framework, we can first build a good-quality average voice from hours of speech and then adapt it to sound like the target speaker from just some minutes of speech data.
• In our work, we investigated the noise tolerance of the HMM adaptation techniques. How does background noise in the speech data affect the Mel-Cepstral component adaptation? In these samples, standard CSMAPLR adaptation has been used, with parameters optimised for noisy source data - no specific noise-robust techniques have been used.

Training data samples * →	Clean speech F: M:	Babble-corrupted speech F: M:	Factory-corrupted speech F: M:	Babble-corrupted speech with noise suppression F: M:	Factory-corrupted speech with noise suppression F: M:
Average voices F: M:	↓	↓	↓	↓	↓
↳	CSMAPLR adaptation with 100 sentences	CSMAPLR adaptation with 100 sentences	CSMAPLR adaptation with 100 sentences	CSMAPLR adaptation with 100 sentences	CSMAPLR adaptation with 100 sentences
	↓	↓	↓	↓	↓
Synthesised samples → (Synthesised MCEP, original F0)	F: M:	F: M:	F: M:	F: M:	F: M:

*) This particular sample is from the test set and has not been used in adaptation

• R. Karhila, U. Remes, and M. Kurimo, "HMM-based speech synthesis adaptation using noisy data: Analysis and evaluation methods," in Proc. ICASSP, 2013
• R. Karhila, U. Remes, and M. Kurimo, "Noise in HMM-based speech synthesis adaptation: Analysis, evaluation methods and experiments," IEEE Journal of Selected Topics in Signal Processing, published on the web 15 August 2013, doi: 10.1109/JSTSP.2013.2278492