uchchaaran

A text-to-speech (TTS) system using concatenative synthesis to generate speech for Devanagari-script (mainly Hindi and Bhojpuri) languages. It combines pre-recorded audio units for natural, clear speech output.

Dataset

The dataset is contained in the data directory.

• data/corpus: Hindi text corpus
• data/devanagari_syllable_dataset: Speech dataset we created for each team member
• data/devanagari_syllable_dataset_split: Segmented speech samples dataset
• data/frequency.json: Frequency data of each segmented syllable like object
• data/syllables.txt: Top 300 syllable like objects sorted by frequency and then by dictionary order used for recording the dataset

Instructions

Installing dependencies

# Initialize and activate a Python virtual environment (optional).
$ python -m venv .venv

# Activate the initialized virtual environment according to your shell (only if using virtual environment).
$ source ./.venv/bin/activate

# Run the following to install the dependencies.
# Or run your system specific command to install.
$ pip install -r requirements.txt

Generating speech

1. Write the text you want it to speak in data/test.txt.
2. Run python stitch.py to save the generate speeches to data/1.wav, data/2.wav, data/3.wav, and data/4.wav using all 4 techniques.

Syllable frequency analysis (already done, DO NOT do it as the new generated order may not match the saved speech samples)

# Generate the frequency.json and syllables.txt files in ./data
$ python main.py

# Generate the split audio units in ./data/devanagari_syllable_dataset_split
$ python audio.py

# Play the audio fragments inside `./data/devanagari_syllable_dataset_split`

Dev Instructions

$ uv venv
$ source ./.venv/bin/activate
$ uv pip compile pyproject.toml -o requirements.txt
$ uv pip sync requirements.txt