jsbeaudry/oswald-large-v3-turbo-m2

🗣️ Haitian Creole (Kreyòl Ayisyen) ASR Model

jsbeaudry/oswald-large-v3-turbo-m2 is a fine-tuned Whisper large-v3-turbo model optimized for Haitian Creole (ht) automatic speech recognition.
The model was trained using Unsloth for efficient low-VRAM fine-tuning.

---

🔍 Model Details

• Base model: openai/whisper-large-v3-turbo
• Task: Automatic Speech Recognition (ASR)
• Language: Haitian Creole (ht)
• Framework: 🤗 Transformers
• Fine-tuning method: Parameter-efficient fine-tuning (PEFT-style)
• Trainable parameters: ~13.1M (≈1.59% of total parameters)
• Output format: Text transcription

---

🧪 Training Summary

Setting	Value
Training samples	9,246
Epochs	1
Total steps	289
Batch size (per device)	4
Gradient accumulation	8
Effective batch size	32
GPUs	1
Precision	FP16
Optimizer	Unsloth default
Checkpointing	Gradient checkpointing enabled

⚠️ use_cache was disabled automatically due to gradient checkpointing.

---

📊 Evaluation Results

Metric: Word Error Rate (WER)
Final validation WER: ~25.96%

WER Progression (selected steps)

Step	Validation WER
5	81.04
25	48.60
50	37.18
100	31.53
150	28.43
200	26.98
250	26.53
280	25.96

The model shows strong convergence within a single epoch, with rapid early gains and stable late-stage improvements.

---

🚀 Usage

Colab (Fast Whisper setup)

https://colab.research.google.com/drive/1_D5KbmhDzRhYHk5xKwP8wEw__IRONWR8?usp=sharing

Transformers (Python)

!pip install transformers

from transformers import pipeline
from IPython.display import Audio, display


asr = pipeline(
    "automatic-speech-recognition",
    model="jsbeaudry/oswald-large-v3-turbo-m2",
    device=0
)


audio_file = "/content/audio_file.wav"
display(Audio(audio_file, rate = 24000))


result = asr(audio_file)
print(result["text"])

Convert for faster_whisper (Python)

!pip install ctranslate2 huggingface_hub faster-whisper

from ctranslate2.converters import TransformersConverter

converter = TransformersConverter(
    "jsbeaudry/oswald-large-v3-turbo-m2",
    low_cpu_mem_usage=True
)
converter.convert("oswald-large-v3-turbo-m2-ct2", quantization="int8")

Use of faster_whisper (Python)

from faster_whisper import WhisperModel

# Load the model, explicitly defining it as a "large-v3" type
# and pointing to the locally converted CTranslate2 model directory.
# This ensures faster_whisper correctly sets n_mels=128.
model = WhisperModel(
    "large-v3", # Specify the model type to ensure correct n_mels (128 for large-v3)
    device="cpu",
    compute_type="int8",
    download_root="/content/oswald-large-v3-turbo-m2-ct2" # Point to the local CT2 model
)

# Transcribe audio
segments, info = model.transcribe(
    "/content/000f7c0f-1661-43d4-819d-91b6a8964b1c.wav",
    language="ht",  # Haitian Creole
    beam_size=5,
    vad_filter=True,  # Recommended for better accuracy
    vad_parameters=dict(min_silence_duration_ms=500)
)

print(f"Detected language '{info.language}' with probability {info.language_probability}")

for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")