Phi-4-mm-inst-asr-turkish-unf

This model is a fine-tuned version of microsoft/Phi-4-multimodal-instruct.

Model Background:
This benchmark evaluates a fine-tuned version of Microsoft's Phi-4-mm-instruct, a multimodal model not originally designed for Turkish ASR. Key points:

1. Initial Limitations:

   • No Turkish ASR support in base model
   • Initial WER 100+%

2. Fine-Tuning Process:

   • Unfroze encoder layers for Turkish adaptation
   • Trained for 1 epoch on Turkish audio-text pairs

3. Current Status:

   • Achieved significant WER reduction (100+% → 9.7% on CommonVoice)*
   • Still under active development for better generalization
   • Results shared as incremental progress documentation

Why This Matters:

• Demonstrates adaptability of multimodal architectures
• Provides baseline for Turkish ASR in resource-constrained scenarios
• Encourages exploration of under-supported languages

• Note on CommonVoice Results:
  • CommonVoice's relatively low WER (9.7%) may benefit from:
    • Potential speaker leakage between splits (same speakers in train/test)
    • Clean audio conditions despite non-professional recordings
    • Short utterance structure (average 4-5 seconds)
  • See below for full context on CommonVoice characteristics in the "Dataset Notes" section.

Benchmark Results

Testing Environment: Google Colab with L4 GPU (24 GB VRAM)

Dataset Notes:

• Finetuning Datasets:

  • commonvoice_17_tr_fixed: Crowd-sourced clean speech (not professional studio recordings) with shuffled splits - potential speaker leakage (same speakers in train/test with different utterances)
  • khanacademy-turkish: Educational lectures with STEM vocabulary
  • yeni-split-0: Noisy real-world recordings

• Benchmark-only Datasets:

  • turkishvoicedataset: Synthetic TTS news (clean but artificial prosody)
  • yeni-split-lq-noisy: Challenging noisy samples with alignment errors

Text Normalization Challenges:
⚠️ Current WER/CER scores may be inflated due to:

1. Lack of standardized Turkish ASR text normalization pipeline
2. Case/punctuation inconsistencies in references
3. Agglutinative language morphology affecting word boundaries

Evaluation Note:
For Turkish ASR benchmarking, I developed a text normalizer to address language-specific scoring challenges. While imperfect, it helps:

• Convert numbers/dates to words
• Standardize compound word formatting
• Reduce punctuation-related mismatches

This preprocessing makes WER/CER calculations slightly fairer compared to raw scoring, though manual verification remains recommended. The tool is actively being refined based on validation set findings.

Performance Factors:

• CommonVoice's relatively low WER (9.7%) likely benefits from:
  • High audio quality despite non-professional speakers
  • Potential speaker familiarity patterns (same speakers in both splits)
  • Short utterance structure (average 4-5 seconds)

Training procedure

finetuning Colab notebook

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.99) and epsilon=1e-07 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Framework versions

• Transformers 4.48.3
• Pytorch 2.5.1+cu124
• Datasets 3.3.2
• Tokenizers 0.21.0