derkar00/mms-tuk-script-latin-adapter

MMS Turkmen ASR Adapter (Latin Script)

This is a lightweight 8MB adapter for Turkmen (tuk-script_latin) Automatic Speech Recognition, extracted from facebook/mms-1b-all.

Instead of downloading the full 3.6GB model, you only need this adapter alongside the base model.

Usage

import librosa
import torch
from transformers import Wav2Vec2ForCTC, AutoProcessor

# Load base model + this adapter
processor = AutoProcessor.from_pretrained("derkar00/mms-tuk-script-latin-adapter")
model = Wav2Vec2ForCTC.from_pretrained("facebook/mms-1b-all")

processor.tokenizer.set_target_lang("tuk-script_latin")
model.load_adapter("tuk-script_latin")

model.eval()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# Load audio (must be 16kHz mono)
audio, _ = librosa.load("your_audio.wav", sr=16000, mono=True)

inputs = processor(audio, sampling_rate=16000, return_tensors="pt").to(device)
with torch.no_grad():
    logits = model(**inputs).logits

predicted_ids = torch.argmax(logits, dim=-1)
print(processor.decode(predicted_ids[0]))

Why only 8MB?

MMS uses an adapter architecture where a single large base model (3.6GB) is shared across 1000+ languages, with small language-specific adapters (~8MB each) plugged in on top.

Details

• Base model: facebook/mms-1b-all
• Language: Turkmen (tuk)
• Script: Latin (tuk-script_latin)
• Task: Automatic Speech Recognition
• Adapter size: 8.4 MB
• License: cc-by-nc-4.0