Audio to SRT Converter with Speaker Diarization

How This Tool Works

Process Overview

1. Audio Upload

   • Upload any audio file (MP3, WAV, M4A, Opus, etc.)
   • File is automatically converted to WAV format for processing

2. Speech-to-Text Transcription

   • Uses OpenAI's Whisper (large-v2 model)
   • Generates accurate word-level timestamps
   • Supports English language

3. Speaker Diarization

   • Uses Pyannote Audio 3.1 for speaker detection
   • Automatically identifies up to 3 different speakers
   • Labels speakers in order of first appearance

4. Text Cleaning & Formatting

   • Removes filler words (um, uh, like, you know, etc.)
   • Splits text into readable sentence blocks
   • Adds speaker labels to each subtitle
   • Generates standard SRT format

Features

• Automatic speaker detection - No manual marking needed
• Custom speaker names - Replace "Speaker 00" with real names
• Clean text - Filler words automatically removed
• Smart formatting - One speaker per subtitle, one sentence per block
• Standard SRT format - Works with all video players and editors
• GPU acceleration - Fast processing on T4 GPU

Tips for Best Results

Before Processing

• Listen to the first minute of your audio to identify speakers
• Note the order speakers appear (first voice = Voice 1, etc.)
• Use clear names for easy identification in subtitles

Audio Quality

• Better audio quality = more accurate transcription
• Minimize background noise for best speaker detection
• Clear speech separation helps diarization accuracy

Speaker Identification

• You don't need to fill in all 3 voices if you have fewer speakers
• If you skip speaker names, output will show "Speaker 00", "Speaker 01", etc.
• Descriptions are just for your reference and don't affect the output

Output Format

Your SRT file will look like this:

1
00:00:01,234 --> 00:00:05,678
(Daniel) Welcome to the podcast.

2
00:00:06,123 --> 00:00:10,456
(Sarah) Thanks for having me.

3
00:00:11,789 --> 00:00:15,234
(Daniel) Let's dive into today's topic.

Each subtitle block includes:

• Subtitle number
• Start and end timestamps (HH:MM:SS,mmm format)
• Speaker name in parentheses
• Cleaned, formatted text

Troubleshooting

"Error: You need to accept the user agreement"

• Visit pyannote/speaker-diarization-3.1
• Click "Agree and access repository"
• Try processing again

"Error: Invalid Hugging Face token"

• Check your token at HF Settings
• Make sure you copied the full token (starts with hf_)
• Generate a new token if needed

Processing takes too long

• Normal processing: 5-15 minutes for typical audio files
• First run may download models (~1-2 GB)
• Longer files (60+ minutes) may take 20-30 minutes

Wrong speaker labels

• Speakers are detected in order of first appearance
• Voice 1 = first person to speak, Voice 2 = second, etc.
• Re-listen to your audio to identify the correct order

Privacy & Security

• Your audio files are processed temporarily and not stored
• Your HF token is only used for this session and never saved
• All processing happens on Hugging Face's secure infrastructure
• Generated SRT files are temporarily stored for download only

Technical Details

Models Used:

• Whisper large-v2 (OpenAI) - Speech-to-text
• Pyannote 3.1 - Speaker diarization

Hardware:

• NVIDIA T4 GPU with CUDA support
• 16GB GPU memory
• Automatic FP16 optimization

Supported Audio Formats: MP3, WAV, M4A, AAC, Opus, FLAC, OGG, WMA, and more

Support

If you encounter issues or have suggestions, please visit the Space's community tab or create an issue.