Number of parameters of tacotron, deep voice, wavenet? I have recently started to explore speech synthesis, and started reading some paper. I have implemented a dummy text to speech synthesis model too, it has around 92 million parameters.
Even though, in computer vision papers, usually the number of parameters of the models are explicitly said, usually in TTS/Speech Recognition I don't see that. What are the number of parameters of the famous TTS models (tacotron, deep voice, wavenet)?
 A: At first, we need to explore basics of DL-based voice synthesis models. A typical end-to-end speech synthesis task is split into 2 subtasks: text -> features representation and features -> speech, the latest one is called vocoder in the literature.
Tacotron2 is an example of model that addresses the first subtask. It maps character embeddings to mel-scale spectrograms and has around 13M parameters with
default setting (source)
.
WaveNet is a vocoder. It could synthesize time-domain waveforms from spectrograms made by Tacotron2. The original paper does not mention details such as number of layers, stacks were used in their experiments. Number of parameters may differ depending on setup. I used this wavenet_vocoder implementation with following hyperparameters to calculate the approximate number of parameters (~3.7M):
out_channels=10 * 3
layers=24
stacks=4
residual_channels=128
gate_channels=256
skip_out_channels=128

from train import build_model
model = build_model()
...
count_parameters(model)
3720932

DeepVoice3 consists of encoder, decoder and converter. At first, it encodes text into per-timestep key and value vectors for an attention-based decoder. The decoder uses these vectors to predict mel-spectrograms that correspond to the audio output. A converter predicts the vocoder parameters for waveform synthesis. (see full explanation).
This implementation contains ~13.3M parameters (calculated as above).
