Use D-vector as speaker embedding

mkotha / WaveRNN

A WaveRNN implementation

MIT License

198 stars 48 forks source link

Use D-vector as speaker embedding #13

Open shaojinding opened 5 years ago

shaojinding commented 5 years ago

Hi,

Does anyone tried to use d-vector as the speaker embedding instead of the one-hot vector? When I was trying to use d-vector, the utilization of latent embedding vectors is bad - the model converges to use only one embedding vector. Any ideas to solve it?

Thanks, Burnie