Method and apparatus for designating a soundalike voice to a target voice from a database of voices
A soundalike system to improve speech synthesis by training a text to speech engine on a voice like the target speakers voice.
1. A speech synthesis system comprising:
a text-to-speech (TTS) system;
a database that stores a collection of voices; and
a processor configured to:
perform training on each voice among the collection of voices stored in the database by building a mathematical model of each voice among the collection of voices stored in the database,
cluster each voice among the collection of voices stored in the database into a voice cluster among a plurality of voice clusters based on similarities between voice characteristics of each voice among the collection of voices stored in the database,
calculate an i-vector for each voice cluster among the plurality of voice clusters, the i-vector for each voice cluster among the plurality of voice clusters representing a mathematical model of each voice cluster among the plurality of voice clusters;
calculate a target i-vector for a target voice,
identify a matching voice cluster among the plurality of voice clusters having an i-vector among the i-vectors for each voice cluster among the plurality of voice clusters that most closely matches the target i-vector for the target voice,
calculate i-vectors of each voice within the matching voice cluster,
identify a soundalike voice that most closely matches the target voice among each voice in the matching voice cluster having an i-vector among the i-vectors for each voice in the matching voice cluster that most closely matches the target i-vector for the target voice, and
build a TTS voice from the soundalike voice for training the TTS.
2. The speech synthesis system of claim 1 , wherein the i-vector of the matching voice cluster has a lowest Euclidean distance from the target i-vector for the target voice among the i-vectors for each voice cluster among the plurality of voice clusters, and
wherein the i-vector of the soundalike voice has a lowest Euclidean distance from the target i-vector for the target voice among the i-vectors for each voice in the matching voice cluster.
3. The speech synthesis system of claim 2 , wherein voice characteristics comprise at least one of speech pitch and speech speed.