Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Fine tuning fastspeech-150k.h5 (v3) on ljspeech for 2k more steps results in poor audio quality.

See original GitHub issue

I’ve downloaded LJSpeech, preprocessed, normalized, extracted durations, and set up for finetuning. I have a dataset for a different voice to do transfer learning, but I was getting poor results, so as a sanity check I tried finetuning on LJSpeech for 2k steps, so total training steps is 152k. This is the result: https://drive.google.com/file/d/1OpcTOx0ozBCVjc8xOHzvPYonyk7I17h5/view?usp=sharing

Result before finetuning: https://drive.google.com/file/d/1tZxQwVsLiSz2VaGkhOPUXYSYyloPFtuU/view?usp=sharing

I was wondering what might cause this gravely noise? It’s much much worse than the 150k model that we started with, which sounds the voice samples linked in the examples page. Durations extracted with models/tacotron2-model-65000.h5.

Issue Analytics

State:
Created 3 years ago
Comments:12 (3 by maintainers)

Top GitHub Comments

1reaction

langfieldcommented, Dec 11, 2020

@dathudeptrai I have tried that and it fixed the finetuning issue. However, with FS2, I’m getting strange behavior when running inference from C++ (tflite). Please see https://github.com/TensorSpeech/TensorFlowTTS/issues/426

0reactions

hertz-pjcommented, Mar 22, 2021

@dathudeptrai I have tried that and it fixed the finetuning issue. However, with FS2, I’m getting strange behavior when running inference from C++ (tflite). Please see #426

@brendanxwhitaker Do you mean that pre-training on fs2 and finetune 2k step can achieve better results?

Top Results From Across the Web

Pedometer-determined step-count guidelines for afterschool ...

This step threshold will provide ASP leaders with an objective, low-cost, easy-to-use tool to monitor progress toward policy-related goals.

The effects of step-count monitoring interventions on physical ...

Step -count monitoring leads to short and long-term step-count increases, with no evidence that either body-worn trackers/smartphone applications ...

How to make your fitness tracker count steps more accurately

Step trackers aren't perfect, but here's what you can do to get the ... the tracker correctly is more likely to result in...

Accurate Step Count with Generalized and Personalized ...

Those studies utilized state-of-the-art deep learning architectures, such as LSTM recurrent network and convolutional neural network, and show promising results ...

Step activity monitor: long-term, continuous recording of ...

Key words: activity, ambulation, functional evaluation, gait, monitor, outcomes assessment, step count. INTRODUCTION. Objective measurements of the effects of ...