Pirnazar/whisper-large-v3-turbo-turkmen-ct2

Whisper large-v3-turbo — Turkmen (faster-whisper)

CTranslate2 FP16 export of Pirnazar/whisper-large-v3-turbo-turkmen, a Turkmen LoRA fine-tune of OpenAI Whisper large-v3-turbo with merged weights. The model downloads automatically on first use.

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel(
    "Pirnazar/whisper-large-v3-turbo-turkmen-ct2",
    device="cpu", compute_type="int8",
)
# With a supported CUDA installation, use device="cuda", compute_type="int8_float16".
segments, info = model.transcribe(
    "audio.wav",
    language="tk",
    task="transcribe",
    beam_size=5,
    without_timestamps=True,
    condition_on_previous_text=False,
)
for segment in segments:
    print(segment.text)

WAV, MP3 and OGG and longer recordings are supported through faster-whisper. Keep without_timestamps=True: the fine-tune was trained without timestamp tokens. Segment times are approximate window positions, not verified word alignment.

Training and evaluation

Base model: OpenAI Whisper large-v3-turbo. Training data: mamed0v/TurkmenSpeech. LoRA rank 8, alpha 16, dropout 0.05, targeting q_proj and v_proj. Exported checkpoint: step 20,565, epoch 3.

The source checkpoint recorded 29.51% WER and 6.59% CER on 500 validation examples during training. These are not independent test metrics or a fresh evaluation of this converted model. Quality on noisy or mixed-language audio is not established. Review important transcripts for errors and hallucinations.

License and attribution

This fine-tuned release is distributed for non-commercial use under CC BY-NC 4.0. The training dataset is published by mamed0v under CC BY-NC 4.0. The original OpenAI Whisper model is MIT-licensed; see LICENSE-WHISPER. Credit OpenAI, mamed0v/TurkmenSpeech and Pirnazar when redistributing this release, and indicate modifications. This is a community fine-tune, not an official OpenAI model release.