ysdede/commonvoice_17_tr_fixed

Improving CommonVoice 17 Turkish Dataset

I recently worked on enhancing the Mozilla CommonVoice 17 Turkish dataset to create a higher quality training set for speech recognition models.
Here's an overview of my process and findings.

Initial Analysis and Split Organization

My first step was analyzing the dataset organization to understand its structure.
Through analysis of filename stems as unique keys, I revealed and documented an important aspect of CommonVoice's design that might not be immediately clear to all users:

• The validated set (113,699 total files) completely contained all samples from:
  • Train split (35,035 files)
  • Test split (11,290 files)
  • Validation split (11,247 files)
• Additionally, the validated set had ~56K unique samples not present in any other split

This design follows CommonVoice's documentation, where dev/test/train are carefully reviewed subsets of the validated data.
However, this structure needs to be clearly understood to avoid potential data leakage when working with the dataset.
For example, using the validated set for training while evaluating on the test split would be problematic since the test data is already included in the validated set.

To create a clean dataset without overlaps, I:

1. Identified all overlapping samples using filename stems as unique keys
2. Removed samples that were already in train/test/validation splits from the validated set
3. Created a clean, non-overlapping validated split with unique samples only

This approach ensures that researchers can either:

• Use the original train/test/dev splits as curated by CommonVoice, OR
• Use my cleaned validated set with their own custom splits

Both approaches are valid, but mixing them could lead to evaluation issues.

Audio Processing and Quality Improvements

Audio Resampling

All audio files were resampled to 16 kHz to:

• Make the dataset directly compatible with Whisper and similar models
• Eliminate the need for runtime resampling during training
• Ensure consistent audio quality across the dataset

Silence Trimming

I processed all audio files to remove unnecessary silence and noise:

• Used Silero VAD with a threshold of 0.6 to detect speech segments
• Trimmed leading and trailing silences
• Removed microphone noise and clicks at clip boundaries

Duration Filtering and Analysis

I analyzed each split separately after trimming silences. Here are the detailed findings per split:

Split	Files Before	Files After	Short Files	Duration Before (hrs)	Duration After (hrs)	Duration Reduction %	Short Files Duration (hrs)	Files Reduction %
Train	11,290	9,651	1,626	13.01	7.34	43.6%	0.37	14.5%
Validation	11,247	8,640	2,609	11.17	6.27	43.9%	0.60	23.2%
Test	35,035	26,501	8,633	35.49	19.84	44.1%	2.00	24.4%
Validated	56,127	46,348	9,991	56.71	32.69	42.4%	2.29	17.4%
Total	113,699	91,140	22,859	116.38	66.14	43.2%	5.26	19.8%

Note: Files with duration shorter than 1.0 seconds were removed from the dataset.

Validation Split Analysis (formerly Eval)

• Original files: 11,247
• Found 2,609 files shorter than 1.0s
• Statistics for short files:
  • Total duration: 26.26 minutes
  • Average duration: 0.83 seconds
  • Shortest file: 0.65 seconds
  • Longest file: 0.97 seconds

Train Split Analysis

• Original files: 35,035
• Found 8,633 files shorter than 1.0s
• Statistics for short files:
  • Total duration: 2.29 hours
  • Average duration: 0.82 seconds
  • Shortest file: 0.08 seconds
  • Longest file: 0.97 seconds

Test Split Analysis

• Original files: 11,290
• Found 1,626 files shorter than 1.0s
• Statistics for short files:
  • Total duration: 56.26 minutes
  • Average duration: 0.85 seconds
  • Shortest file: 0.65 seconds
  • Longest file: 0.97 seconds

Validated Split Analysis

• Original files: 56,127
• Found 9,991 files shorter than 1.0s
• Statistics for short files:
  • Total duration: 36.26 minutes
  • Average duration: 0.83 seconds
  • Shortest file: 0.65 seconds
  • Longest file: 0.97 seconds

All short clips were removed from the dataset to ensure consistent quality. The final dataset maintains only clips longer than 1.0 seconds, with average durations between 2.54-2.69 seconds across splits.

Final Split Statistics

The cleaned dataset was organized into:

• Train: 26,501 files (19.84 hours, avg duration: 2.69s, min: 1.04s, max: 9.58s)
• Test: 9,650 files (7.33 hours, avg duration: 2.74s, min: 1.08s, max: 9.29s)
• Validation: 8,639 files (6.27 hours, avg duration: 2.61s, min: 1.04s, max: 9.18s)
• Validated: 46,345 files (32.69 hours, avg duration: 2.54s, min: 1.04s, max: 9.07s)

Final Dataset Split Metrics

Split	Files	Duration (hours)	Avg Duration (s)	Min Duration (s)	Max Duration (s)
TRAIN	26501	19.84	2.69	1.04	9.58
TEST	9650	7.33	2.74	1.08	9.29
VALIDATION	8639	6.27	2.61	1.04	9.18
VALIDATED	46345	32.69	2.54	1.04	9.07

Total files processed: 91,135 Valid entries created: 91,135 Files skipped: 0 Total dataset duration: 66.13 hours Average duration across all splits: 2.61 seconds

The dataset was processed in the following order:

1. Train split (26,501 files)
2. Test split (9,650 files)
3. Validation split (8,639 files) - Note: Also known as "eval" split in some CommonVoice versions
4. Validated split (46,348 files)

Note: The validation split (sometimes referred to as "eval" split in CommonVoice documentation) serves the same purpose - it's a held-out set for model validation during training.
We've standardized the naming to "validation" throughout this documentation for consistency with common machine learning terminology.

One text file in the validated split was flagged for being too short (2 characters), but was still included in the final dataset.

The processed dataset was saved as 'commonvoice_17_tr_fixed'.

Text Processing and Standardization

Character Set Optimization

• Created a comprehensive charset from all text labels
• Simplified the character set by:
  • Standardizing quotation marks
  • Removing infrequently used special characters

Text Quality Improvements

• Generated word frequency metrics to identify potential issues
• Corrected common Turkish typos and grammar errors
• Standardized punctuation and spacing

Results

The final dataset shows significant improvements:

• Removed unnecessary silence and noise from audio
• Consistent audio durations above 1.0 seconds
• Standardized text with corrected Turkish grammar and typography
• Maintained original metadata (age, upvotes, etc.)

These improvements make the dataset more suitable for training speech recognition models while maintaining the diversity and richness of the original CommonVoice collection.

Tools Used

This dataset processing work was completed using ASRTK (Automatic Speech Recognition Toolkit), an open-source Python toolkit designed to streamline the development and enhancement of ASR systems. ASRTK provides utilities for:

• Audio processing with advanced splitting and resampling capabilities
• Text normalization and cleaning
• Forced alignment using Silero VAD models
• Efficient batch processing with multi-threading support

The toolkit is available under the MIT license and welcomes contributions from the community.