ChineseTtsTflite

Chinese TTS TF Lite

introduce

The TTS engine developed using Kotlin + JetPack Compose + Tensorflow Lite can be used completely offline.

Two optional models: FastSpeech and Tacotron, both of which are from TensorFlowTTS

The method of text to pinyin comes from: TensorflowTTS_chinese

Because it is real-time inference output audio, there are certain requirements for device performance.

Among them, FastSpeech is faster, but the generated audio has poor anthropomorphism effect and can be used on ordinary mid-range or above mobile phones.

Tacotron has high performance requirements. Although the overall effect is better, it is very slow, so it is currently of little practical value and is only for testing.

Manual compilation instructions

1. Android Studio 2021.2.1
2. From the following model download address, download models-tflite.7z, and unzip the following 4 files to the corresponding directory
3. From the aar download address below, download the cropped TensorFlow Lite aar, and put 2 aar files into app/libs

 ├─app/src/main/assets
│      baker_mapper.json
│      fastspeech2_quan.tflite
│      mb_melgan.tflite
│      tacotron2_quan.tflite

3. Compilation

./gradlew assembleRelease

Model download

Download address

• models-tf.7z: original TensorflowTTS model, generally used on PC side
• models-tflite.7z: The converted TFLite model, generally used in mobile terminals

Model View

netron

Crop TensorFlow Lite

Reduce the TensorFlow Lite binary file size

Download aar after cropping

Comparison of arr file sizes before and after cropping:

• tensorflow-lite-2.8.0.aar: 5.4MB -> 3.7MB (68.5%)
• tensorflow-lite-select-tf-ops-2.8.0.aar: 109.6MB -> 14.8MB (13.5%)

References

TensorFlowTTS

TensorflowTTS_chinese

Tensorflow Guide

Reference tf test

Reference tflite conversion

Google Pico TTS Source

Model Test & Conversion

Convert environment configuration

• Ubuntu: 20.04 LTS
• Python: 3.8

$ git clone https://github.com/TensorSpeech/TensorFlowTTS.git
$ cd TensorFlowTTS
$ pip install .
$ pip install git+https://github.com/repodiac/german_transliterate.git

Decompress models-tf.7z

TF model testing

$ cd models-tf
$ python test-h5.py

TF model to TFLite

$ python convert-tflite.py

Screenshot of APP interface