Examples of synthesized utterances

Effect of HTS-training database size on synthesis



Example utterances of this page are generated using HMM-based synthesis trained with a Finnish speech database. Voice hanna_hts is trained using the full database of 1.3 hours. Voices hanna_hts_100 and hanna_hts_30 are trained using smaller databases of 100 and 30 sentences, respectively. Construction of the smaller databases is done by selecting greedily sentences with good triphone coverage from the full database. Both of the smaller databases have been ensured to have a full coverage for the phones (for both short and long quanties, as well as for diphthongs) occurring in the full database.

Sampling rate of the utterances is 16 kHz. You can ensure that you browser handles it correctly with this test sentence . Utterance is recorded and should not contain distortion.


Prosody

Utterances generated with voices hanna_hts and hanna_hts_100.

hanna_hts hanna_hts_100
hanna_hts hanna_hts_100
hanna_hts hanna_hts_100
hanna_hts hanna_hts_100

Intelligibility

Utterances with a nonsense word as a subject ('Suuri .... asuu mökissä', 'A big .... lives in a cottage') generated using voices hanna_hts, hanna_hts_100, hanna_hts_30.

hanna_hts hanna_hts_100 hanna_hts_30
hanna_hts hanna_hts_100 hanna_hts_30
hanna_hts hanna_hts_100 hanna_hts_30

last modified: 2008-04-15 HS