Update README.md

README.md

@@ -46,9 +46,9 @@ pipeline_tag: automatic-speech-recognition
 library_name: transformers
 ---
 
-# Whisper-Large-v3 Dutch - High + Mixed Quality Filtered Synthetic Data
+# Whisper-Large-v3 Dutch - High-Quality Filtered Synthetic Data
 
-This model is a fine-tuned version of [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) for Dutch automatic speech recognition (ASR). It was trained on Common Voice 17.0 Dutch combined with **WAVe-filtered synthetic speech data** including both high-quality (q ≥ 0.8) and medium-quality (0.5 ≤ q < 0.8) samples, using threshold q ≥ 0.5.
+This model is a fine-tuned version of [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3) for Dutch automatic speech recognition (ASR). It was trained on Common Voice 17.0 Dutch combined with **WAVe-filtered high-quality synthetic speech data only** using a strict threshold (q ≥ 0.8).
 
 ## Introduction
 
@@ -60,19 +60,19 @@ The training data combines real speech from Common Voice 17.0 with synthetic spe
 
 2. **Speech Synthesis**: Each transcript was converted to audio using OpenAI's TTS-1 model with 9 different voice variants (alloy, ash, coral, echo, fable, nova, onyx, sage, shimmer), producing 34,898 synthetic samples.
 
-3. **Quality Filtering with WAVe**: Raw synthetic speech often contains defects such as mispronunciations, omitted words, or prosodic anomalies. To address this, we applied **WAVe (Word-Aligned Verification)**, a model that assesses audio-text alignment at the word level rather than the sentence level. WAVe uses multi-head attention to align each word to its corresponding audio frames and assigns per-word confidence scores via a GLU-based scorer. For this model, we retained samples scoring above the balanced threshold (q ≥ 0.5), including both high-quality (q ≥ 0.8) and medium-quality (0.5 ≤ q < 0.8) samples, resulting in 30,182 filtered synthetic samples.
+3. **Quality Filtering with WAVe**: Raw synthetic speech often contains defects such as mispronunciations, omitted words, or prosodic anomalies. To address this, we applied **WAVe (Word-Aligned Verification)**, a model that assesses audio-text alignment at the word level rather than the sentence level. WAVe uses multi-head attention to align each word to its corresponding audio frames and assigns per-word confidence scores via a GLU-based scorer. For this model, only samples scoring above the strict threshold (q ≥ 0.8) were retained, resulting in 10,555 high-quality synthetic samples.
 
 ### How the Model Was Created
 
 The model was fine-tuned from `openai/whisper-large-v3` using the Hugging Face Transformers library with the following approach:
 
-1. **Mixed Training**: Combined 34,952 real speech samples from Common Voice 17.0 Dutch with 30,182 WAVe-filtered synthetic samples (high + medium quality, q ≥ 0.5), for 65,134 total samples.
+1. **Mixed Training**: Combined 34,952 real speech samples from Common Voice 17.0 Dutch with 10,555 strictly WAVe-filtered high-quality synthetic samples (45,507 total).
 
 2. **Optimization**: Trained for 5 epochs with a learning rate of 5e-6, global batch size of 256, and BF16 precision on an NVIDIA H200 GPU.
 
 3. **Checkpoint Selection**: The best checkpoint was selected based on validation loss, occurring at step 350 with a validation loss of 0.0552.
 
-This balanced filtering approach achieves **7% reduction in training steps** compared to using all synthetic data, while filtering out only the lowest-quality 13.5%.
+This high-quality filtering approach achieves **35% reduction in training steps** compared to using all synthetic data, while maintaining excellent ASR performance.
 
 ## Model Details
 
@@ -82,8 +82,8 @@ This balanced filtering approach achieves **7% reduction in training steps** com
 | **Language** | Dutch (nl) |
 | **Task** | Automatic Speech Recognition (transcribe) |
 | **Parameters** | 1550M |
-| **Training Data** | Common Voice 17.0 + High + Mixed Quality Synthetic (q ≥ 0.5) |
-| **Total Training Samples** | 65,134 |
+| **Training Data** | Common Voice 17.0 + High-Quality Synthetic (q ≥ 0.8) |
+| **Total Training Samples** | 45,507 |
 | **Sampling Rate** | 16kHz |
 
 ## Evaluation Results
@@ -110,11 +110,11 @@ This balanced filtering approach achieves **7% reduction in training steps** com
 
 ### Key Performance Highlights
 
-- **Efficient training**: Only 890 max steps (7% fewer than unfiltered)
+- **Most efficient training**: Only 890 max steps (35% fewer than unfiltered)
 - **Best validation loss** (0.0520) among all Whisper-Large-v3 Dutch configurations
-- **Strong cross-domain performance**: 20.29% MLS WER (9.5% relative improvement vs baseline)
-- **Competitive in-domain**: 4.43% Test WER on Common Voice
-- **Balanced filtering**: Retains 86.5% of synthetic data (high + medium quality)
+- **Competitive in-domain performance**: 4.43% Test WER on Common Voice
+- **9.5% relative improvement** on MLS benchmark vs baseline (20.29% vs 22.43%)
+- **Best quality-to-compute ratio**: Strong results with only top-tier synthetic data (30.2%)
 
 ## Training Data
 
@@ -123,8 +123,8 @@ This balanced filtering approach achieves **7% reduction in training steps** com
 | Source | Samples | Description |
 |--------|---------|-------------|
 | [Common Voice 17.0 Dutch](https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0) | 34,952 | Real speech from Mozilla's crowdsourced dataset |
-| [Synthetic Transcript NL](https://huggingface.co/datasets/yuriyvnv/synthetic_transcript_nl) (q ≥ 0.5) | 30,182 | WAVe-filtered TTS audio (high + medium quality) |
-| **Total** | **65,134** | |
+| [Synthetic Transcript NL](https://huggingface.co/datasets/yuriyvnv/synthetic_transcript_nl) (q ≥ 0.8) | 10,555 | Strictly WAVe-filtered TTS audio (high quality only) |
+| **Total** | **45,507** | |
 
 ### Synthetic Data Generation Pipeline
 
@@ -132,17 +132,17 @@ The synthetic dataset ([yuriyvnv/synthetic_transcript_nl](https://huggingface.co
 
 1. **Transcript Generation**: GPT-4o-mini, matching Common Voice word count distribution
 2. **Speech Synthesis**: OpenAI TTS-1 model with 9 voice variants (alloy, ash, coral, echo, fable, nova, onyx, sage, shimmer)
-3. **Quality Filtering**: WAVe model with balanced threshold q ≥ 0.5 (includes high + medium quality)
+3. **Quality Filtering**: WAVe model with strict threshold q ≥ 0.8 (high quality only)
 
 ### WAVe Quality Distribution (Dutch Synthetic Data)
 
 | Quality Level | Samples | Percentage | Used in This Model |
 |--------------|---------|------------|-------------------|
 | High (q ≥ 0.8) | 10,555 | 30.2% | ✓ |
-| Medium (0.5 ≤ q < 0.8) | 19,627 | 56.2% | ✓ |
+| Medium (0.5 ≤ q < 0.8) | 19,627 | 56.2% | ✗ |
 | Low (q < 0.5) | 4,716 | 13.5% | ✗ |
 
-This threshold retains 86.5% of the synthetic dataset (high + medium quality), filtering only the lowest-quality samples.
+This strict threshold retains only the top 30.2% of synthetic samples, prioritizing quality over quantity for maximum training efficiency.
 
 ## Training Procedure
 
@@ -228,18 +228,18 @@ This model leverages **WAVe (Word-Aligned Verification)**, a word-level quality
 - Detects localized synthesis errors (mispronunciations, omitted words, prosodic anomalies)
 - Achieves **6.5% improvement** over sentence-level filtering methods
 
-The balanced threshold (q ≥ 0.5) retains both high-quality and medium-quality samples (86.5% of synthetic data), filtering only the lowest 13.5% while maintaining data volume for robust training.
+The strict threshold (q ≥ 0.8) retains only the top 30.2% of synthetic samples, prioritizing quality over quantity for maximum training efficiency.
 
 ## When to Use This Model
 
 This model is ideal when:
-- **Best validation performance required**: Achieves lowest validation loss (0.0520) among all Large-v3 Dutch variants
-- **Balanced approach needed**: Includes 86.5% of synthetic data (high + medium quality)
-- **Compute efficiency matters**: 7% fewer training steps than unfiltered
-- **Strong cross-domain performance desired**: 9.5% relative improvement on MLS vs baseline
+- **Compute resources are limited**: 35% fewer training steps than unfiltered approaches
+- **Quick fine-tuning is needed**: Smaller dataset (45,507 samples) enables faster iteration
+- **Best validation performance required**: Achieves lowest validation loss (0.0520)
+- **Quality over quantity**: Only top-tier synthetic data (30.2%) for clean training signal
 
 Consider other variants based on your needs:
-- [whisper-large-v3-mixed-cv-nl](https://huggingface.co/yuriyvnv/whisper-large-v3-mixed-cv-nl): Same data as this model (alternative naming)
+- [whisper-large-v3-mixed-cv-nl](https://huggingface.co/yuriyvnv/whisper-large-v3-mixed-cv-nl): Better cross-domain performance with more data
 - [whisper-large-v3-cv-fully-synthetic-nl](https://huggingface.co/yuriyvnv/whisper-large-v3-cv-fully-synthetic-nl): Best cross-domain generalization (17.02% MLS)
 
 ## Limitations