prof-freakenstein/whisper-librispeech-finetuned

Fine-tuned openai/whisper-small on the LibriSpeech ASR dataset (clean subset, train.100 split).

Model Details

Property	Value
Base model	openai/whisper-small
Language	English
Task	Automatic Speech Recognition (transcribe)
Dataset	LibriSpeech ASR – clean / train.100
Training epochs	3
Batch size	16 (per device) × 2 grad-accum = 32 effective
Learning rate	1e-05
Warmup steps	500
Precision	FP16
Hardware	NVIDIA L4, Intel Xeon

Training Configuration

# =============================================================================
#  Whisper Fine-Tuning Configuration – LibreSpeech on NVIDIA L4 / Intel Xeon
# =============================================================================

# ---------------------------------------------------------------------------
# Model
# ---------------------------------------------------------------------------
model:
  name: "openai/whisper-small"      # Options: whisper-tiny | whisper-base | whisper-small | whisper-medium | whisper-large-v3
  language: "english"
  task: "transcribe"
  # Set to false to freeze encoder weights and only train the decoder
  freeze_encoder: false
  # Set to true to freeze encoder weights (faster training, lower VRAM)
  # freeze_feature_extractor is not used; configure freezing via freeze_encoder above

# ---------------------------------------------------------------------------
# Dataset  – LibreSpeech via HuggingFace datasets hub
# ---------------------------------------------------------------------------
dataset:
  name: "librispeech_asr"
  config_name: "clean"              # "clean" | "other"
  train_split: "train.100"          # train.100 | train.360 | train.500
  validation_split: "validation"
  test_split: "test"
  streaming: false                  # Set true to avoid downloading full dataset
  num_proc: 8                       # Parallel CPU workers for preprocessing (Xeon friendly)
  cache_dir: "./data/cache"

# ---------------------------------------------------------------------------
# Feature extraction
# ---------------------------------------------------------------------------
feature_extraction:
  sampling_rate: 16000
  max_input_length_seconds: 30.0

# ---------------------------------------------------------------------------
# Training  – tuned for NVIDIA L4 (24 GB VRAM)
# ---------------------------------------------------------------------------
training:
  output_dir: "./outputs/whisper-librispeech"
  num_train_epochs: 3
  per_device_train_batch_size: 16
  per_device_eval_batch_size: 8
  gradient_accumulation_steps: 2     # Effective batch = 32 per GPU
  learning_rate: 1.0e-5
  warmup_steps: 500
  weight_decay: 0.01
  lr_scheduler_type: "linear"
  fp16: true                         # L4 supports FP16; set bf16: true if preferred
  bf16: false
  dataloader_num_workers: 4          # Intel Xeon – use multiple cores
  save_strategy: "epoch"
  evaluation_strategy: "epoch"
  logging_steps: 25
  save_total_limit: 3
  load_best_model_at_end: true
  metric_for_best_model: "wer"
  greater_is_better: false
  predict_with_generate: true
  generation_max_length: 225
  report_to: "tensorboard"
  seed: 42

# ---------------------------------------------------------------------------
# Hardware hints
# ---------------------------------------------------------------------------
hardware:
  gpu: "NVIDIA L4"
  vram_gb: 24
  cpu: "Intel Xeon"
  # Use torch.compile for extra throughput (requires PyTorch >= 2.0)
  use_torch_compile: false

Usage

from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="prof-freakenstein/whisper-librispeech-finetuned",
    chunk_length_s=30,
    stride_length_s=5,
)
result = asr("path/to/audio.wav")
print(result["text"])

Or with the WhisperProcessor / WhisperForConditionalGeneration API:

import torch
from transformers import WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("prof-freakenstein/whisper-librispeech-finetuned")
model = WhisperForConditionalGeneration.from_pretrained("prof-freakenstein/whisper-librispeech-finetuned")
model.eval()

# load your audio as a 16 kHz numpy array `audio_array` …
inputs = processor(audio_array, sampling_rate=16_000, return_tensors="pt")
with torch.no_grad():
    predicted_ids = model.generate(inputs["input_features"])
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(transcription[0])

Training Script

This model was trained using the transcription-as-thought pipeline:

python scripts/train.py --config config/training_config.yaml

Limitations

• Optimised for English speech; accuracy on other languages may vary.
• Trained on read/narrated speech (LibriSpeech); performance on conversational or noisy audio may be lower than on clean recordings.

License

This model is released under the Apache 2.0 licence, consistent with the base Whisper model.