Whisper large-v3-turbo — Turkmen

Turkmen speech recognition model shared by Pirnazar. This is a standalone Whisper large-v3-turbo checkpoint with a trained LoRA adapter merged into the weights. No separate adapter or base-model download is needed.

Run with Transformers

pip install "transformers==4.57.6" torch soundfile scipy

Example for a WAV file of up to 30 seconds:

import math
import soundfile as sf
import torch
from scipy.signal import resample_poly
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq

repo = "Pirnazar/whisper-large-v3-turbo-turkmen"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if device == "cuda" else torch.float32
processor = AutoProcessor.from_pretrained(repo)
model = AutoModelForSpeechSeq2Seq.from_pretrained(repo, torch_dtype=dtype).to(device)
model.eval()
audio, sr = sf.read("audio.wav", dtype="float32")
if audio.ndim == 2:
    audio = audio.mean(axis=1)
if sr != 16000:
    divisor = math.gcd(sr, 16000)
    audio = resample_poly(audio, 16000 // divisor, sr // divisor)
if len(audio) > 30 * 16000:
    raise ValueError("Use the faster-whisper version below for longer audio.")
inputs = processor(audio, sampling_rate=16000, return_tensors="pt", return_attention_mask=True)
with torch.inference_mode():
    ids = model.generate(
        inputs.input_features.to(device=device, dtype=dtype),
        attention_mask=inputs.attention_mask.to(device),
        language="tk", task="transcribe", return_timestamps=False,
    )
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

For longer recordings, MP3/OGG input and faster inference, use the CTranslate2 version.

Training and evaluation

• Base: OpenAI Whisper large-v3-turbo.
• Data: mamed0v/TurkmenSpeech.
• LoRA: rank 8, alpha 16, dropout 0.05, targeting q_proj and v_proj.
• Exported checkpoint: step 20,565, epoch 3.
• Recorded validation results on 500 examples: WER 29.51%, CER 6.59%.

These are the recorded training-time validation metrics, not an independent test-set evaluation or a fresh benchmark of this export. Performance on other speakers, noise and mixed-language audio has not been established.

Limitations

Use language="tk" and transcription mode. Training did not use timestamp tokens: disable timestamp generation. Segment times in the CTranslate2 version are approximate window positions, not verified word alignment. Recognition can contain errors and hallucinations; review important transcripts.

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.