Dataset Creator for TTS Training

A Python tool for creating text-to-speech (TTS) training datasets by recording audio for text phrases. This tool provides an interactive command-line interface for recording audio samples with automatic silence removal, audio validation, and progress tracking.

Features

• 🎤 High-quality audio recording using sounddevice (22.05 kHz sample rate)
• 📝 Flexible input formats - supports both wavfile|text and plain text formats
• ✂️ Automatic silence removal with fade in/out to prevent clicks
• ✅ Audio validation - automatic testing of recorded audio quality
• 📊 Progress tracking - real-time display of recording progress
• 💾 Auto-save - saves progress on interruption (Ctrl+C)
• 🔄 Re-record option - easily re-record any phrase
• ⏭️ Skip phrases - skip unwanted phrases without recording
• 🚪 Escape mode - save current recording and exit gracefully
• 🎯 Random filenames - generates unique 8-character filenames for audio files

Requirements

Python Dependencies

• Python 3.6+
• librosa - Audio processing
• soundfile - Audio file I/O
• sounddevice - Audio recording
• numpy - Numerical operations

System Requirements

• A working microphone/audio input device
• Terminal/command-line interface
• Sufficient disk space for audio files

Installation

1. Install the required Python packages:

pip install librosa soundfile sounddevice numpy

2. Make the script executable (optional):

chmod +x dataset_creator_final.py

Configuration

Edit the CONFIG dictionary in the main() function to set your paths:

CONFIG = {
    'input_text': 't1.txt',      # Input text file
    'output_text': 't.txt',      # Output dataset file
    'wave_path': 'audio/',       # Directory to save WAV files
}

Usage

Basic Usage

python dataset_creator_final.py

Or if made executable:

./dataset_creator_final.py

Workflow

1. Prepare input file: Create a text file with phrases to record (see Input Formats below)
2. Run the script: Execute dataset_creator_final.py
3. Record phrases: For each phrase:
   • Press Enter to start recording
   • Speak the phrase
   • Press Enter to stop recording
   • Choose an action (see Recording Controls below)
4. Review output: The tool automatically validates audio and saves the dataset

Input File Formats

The tool automatically detects and supports two input formats:

Format 1: Plain Text (one phrase per line)

Hello world
This is a test
How are you today?

Format 2: wavfile|text Format

dummy.wav|Hello world
dummy.wav|This is a test
dummy.wav|How are you today?

Note: In wavfile|text format, the tool extracts only the text portion (after the |). The wavfile name before | is ignored but can be useful for reference.

Output Format

The output dataset file contains one entry per line in the format:

audio/abc12345.wav|Hello world
audio/def67890.wav|This is a test
audio/ghi11111.wav|How are you today?

Where:

• audio/ is the wave_path directory
• abc12345.wav is a randomly generated 8-character filename
• | separates the audio file path from the text
• The text is the original phrase

Recording Controls

After recording a phrase, you'll be prompted with these options:

• y or yes - Save this recording and move to next phrase
• n or no - Record again (re-record the same phrase)
• s or skip - Skip this phrase (don't save, move to next)
• e or escape - Save current recording and exit (saves all progress)

Keyboard Interrupt

Press Ctrl+C at any time to:

• Save all recorded files to the output dataset
• Exit gracefully
• Preserve all progress made so far

Audio Processing

Automatic Features

1. Silence Removal: Automatically trims silence from the beginning and end of recordings

   • Threshold: 0.005 (configurable)
   • Padding: 50ms before/after detected audio

2. Click Prevention: Applies fade in/out (10ms) to prevent audio clicks

3. Format Standardization:

   • Sample rate: 22.05 kHz
   • Format: WAV (PCM_16)
   • Channels: Mono

Audio Validation

After recording, the tool automatically:

• Checks recording duration
• Validates audio amplitude
• Detects potential clipping
• Measures energy levels
• Warns about very short/long recordings

Example Session

🎯 Final Dataset Creator for TTS Training
==================================================
Input file: t1.txt
Output file: t.txt
Audio directory: audio/
==================================================

📝 Detected format: plain text (one per line)
📊 Loaded 3 phrases from plain_text format

📝 Progress: 1/3 | 🎵 Saved WAVs: 0

============================================================
Hello world
============================================================
Press Enter to start recording...
🎤 Recording... Press Enter to stop.
(Recording...)
Press Enter to stop recording...
✅ Recording completed successfully
⏱️  Recording duration: 2.34 seconds

Options:
  y - Save this recording
  n - Record again
  s - Skip this text (don't save, move to next)
  e - Escape (save all data and exit)

What would you like to do? (y/n/s/e): y
✅ Saved as: abc12345.wav (2.34s)
✅ Added to dataset: abc12345.wav
   Total recorded so far: 1

File Structure

After running the tool, you'll have:

project/
├── dataset_creator_final.py
├── t1.txt                    # Input file
├── t.txt                     # Output dataset file
└── audio/                    # Audio files directory
    ├── abc12345.wav
    ├── def67890.wav
    └── ...

Troubleshooting

No Audio Recorded

• Check microphone permissions: Ensure your system allows microphone access
• Check audio device: Verify your microphone is connected and working
• Test with system audio tools: Try recording with other applications first

Audio Quality Issues

• Low volume: Check microphone input levels in system settings
• Clipping detected: Reduce microphone input gain
• Very short duration: Ensure you're speaking clearly and loudly enough

File Not Found Errors

• Input file: Ensure t1.txt (or your configured input file) exists
• Output directory: The tool creates audio/ automatically, but ensure write permissions

Import Errors

If you get import errors, install missing packages:

pip install librosa soundfile sounddevice numpy

Technical Details

• Sample Rate: 22.05 kHz (configurable in FinalAudioRecorder.__init__)
• Audio Format: WAV, PCM_16, Mono
• Silence Threshold: 0.005 (configurable in remove_silence())
• Fade Duration: 10ms
• Padding: 50ms before/after detected audio

License

This tool is part of the shramVoice project. Modify and use as needed for your TTS training workflow.

Contributing

To modify the tool:

• Edit the CONFIG dictionary in main() for paths
• Adjust sample_rate in FinalAudioRecorder.__init__() for different sample rates
• Modify remove_silence() threshold for different silence detection sensitivity