Sdab

Khmer Automatic Speech Recognition (Whisper + Wav2Vec2)

Sdab is a lightweight helper around Hugging Face ASR models with a focus on Khmer language. It can load sequence-to-sequence Whisper checkpoints (default) or CTC-style Wav2Vec2 models, convert audio to the expected format, and return a transcription in a single call.

• License: Apache-2.0
• Default Whisper model: metythorn/whisper-large-v3
• Optional turbo Whisper model: metythorn/whisper-large-v3-turbo
• Example Wav2Vec2 model: metythorn/wav2vec2-xls-r-300m

Features

• 🔁 Automatically detects Whisper vs Wav2Vec2 when you pass a Hugging Face repo ID or local directory.
• 🎧 Handles loading, mono conversion, and resampling to 16 kHz with torchaudio.
• ⚙️ Lets you pick CPU/GPU device and numerical precision to match your hardware.
• 🧪 Includes a sample audio clip for quick testing.

Installation

It is highly recommended to work inside a virtual environment.

python -m pip install --upgrade pip
pip install torch torchaudio transformers soundfile
pip install sdab

GPU users should install the right torch/torchaudio binaries for their CUDA version as described on https://pytorch.org/get-started/locally/.

To install from source:

git clone https://github.com/MetythornPenn/sdab.git
cd sdab
pip install -e .

Quick Start

Download the bundled sample audio (Khmer speech, 16 kHz WAV):

wget -O audio.wav https://raw.githubusercontent.com/MetythornPenn/sdab/main/sample/audio.wav

Whisper (default)

from sdab import Sdab

sd = Sdab("audio.wav")  # defaults to metythorn/whisper-large-v3
print(sd.transcribe())

Explicit Whisper model

from sdab import Sdab

sd = Sdab(
    "audio.wav",
    model_name="metythorn/whisper-large-v3",
    device="cuda:0",         # or "cpu"
)
print(sd.transcribe())

Need the faster turbo checkpoint? Provide it explicitly:

sd = Sdab("audio.wav", model_name="metythorn/whisper-large-v3-turbo")
print(sd.transcribe())

Wav2Vec2 / CTC model

from sdab import Sdab

sd = Sdab(
    "audio.wav",
    model_name="metythorn/wav2vec2-xls-r-300m",
    model_type="wav2vec2",  # optional; Sdab will infer from the model name
)
print(sd.transcribe())

Important parameters

• file_path: Path to your WAV/FLAC/etc. file.
• model_name: Hugging Face repo ID or local directory with the pretrained model.
• model_type: Force "whisper" or "wav2vec2" if autodetect is not correct.
• device: "cpu" or any PyTorch device string (for example "cuda:0").
• torch_dtype: Override the dtype (defaults to float32 on CPU and float16 on CUDA).

Tips

• Whisper expects mono 16 kHz input; Sdab automatically downsamples and squeezes channels.
• Models are downloaded from Hugging Face the first time you reference them. Keep an eye on cache size in ~/.cache/huggingface.
• For long recordings consider chunking/streaming outside of Sdab to stay within GPU memory.
• Results are returned from sd.transcribe() directly; the class no longer stores a separate sd.result.
• Errors while loading a model are wrapped in a helpful RuntimeError with the model name.

References

• Inspired by Bong Vitou Phy and the accompanying Techcast episode.
• Khmer word segmentation libraries from SeangHay: khmercut and khmersegment.
• Whisper: paper | Hugging Face models.
• Wav2Vec2 paper and resources from Facebook AI Research: fairseq examples.