Noise at the end of produced wave file

[iarspider]

🐛 Bug

Noise at the end of produced wave file

To Reproduce

Steps to reproduce the behavior:

1. Install torch, numpy
2. Run the example for Habrahabr article:

import os
import torch

device = torch.device('cpu')
torch.set_num_threads(4)
local_file = 'model.pt'

if not os.path.isfile(local_file):
    torch.hub.download_url_to_file('https://models.silero.ai/models/tts/ru/v2_kseniya.pt',
                                   local_file)  

model = torch.package.PackageImporter(local_file).load_pickle("tts_models", "model")
model.to(device)

example_batch = ['В недрах тундры выдры в г+етрах т+ырят в вёдра ядра кедров.',
                 'Котики - это жидкость!',
                 'М+ама М+илу м+ыла с м+ылом.']
sample_rate = 16000

audio_paths = model.save_wav(texts=example_batch,
                             sample_rate=sample_rate)

3. The second and third audio files are 12 seconds long (instead of ~1 sec), and are "padded" with noise
   

Expected behavior

No padding with noise

Environment

Please copy and paste the output from this
environment collection script
(or fill out the checklist below manually).

Collecting environment information...
PyTorch version: 1.9.0+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
ROCM used to build PyTorch: N/A

OS: Microsoft Windows 10 Pro
GCC version: Could not collect
Clang version: Could not collect
CMake version: Could not collect
Libc version: N/A

Python version: 3.8 (64-bit runtime)
Python platform: Windows-10-10.0.18362-SP0
Is CUDA available: False
CUDA runtime version: Could not collect
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 3080
Nvidia driver version: 466.77
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A

Versions of relevant libraries:
[pip3] numpy==1.20.3
[pip3] torch==1.9.0
[conda] Could not collect

Additional context

Thanks a lot for creating this!

[snakers4]

This is a known tacotron bug
Happens due to batching

Batch size 1 fixes this (i.e. just feed 1 phrase)

We basically have 2 options how to fix this more properly

• Downstream with a VAD (please use silero-vad or WebRTC)
• By looking at the alignment map of tacotron (we did not implement this yet)

[snakers4]

in future though we will stop using tacotron, so the maybe just using batch_size = 1 or just applying a VAD is good temporary solution (especially when you run on CPU)