Wav2Vec2 Large XLSR Turkish ONNX

ONNX export of m3hrdadfi/wav2vec2-large-xlsr-turkish for asr.js.

Model details

• Architecture: Wav2Vec2 Large XLSR-53 (24 transformer layers, 16 heads, 1024 hidden)
• Parameters: ~317M
• Input: Raw 16 kHz mono PCM waveform [1, samples] float32
• Output: CTC logits [1, frames, 40] float32
• Output stride: 320 samples (50 fps at 16 kHz)
• Vocabulary: 40 character-level tokens for Turkish (includes ç, ğ, ı, ö, ş, ü)
• Languages: Turkish (tr)
• Fine-tuned on: Common Voice Turkish
• ONNX opset: 18
• External data: Yes (1.2 GB model weights in .data file)

Usage with asr.js

import { loadSpeechModel } from '@asrjs/speech-recognition';
import { createWav2Vec2AlignerFromLogits } from '@asrjs/speech-recognition/alignment';

const model = await loadSpeechModel('m3hrdadfi/wav2vec2-large-xlsr-turkish', {
  useManifestSources: true,
});

// ASR
const transcript = await model.transcribe(turkishAudio);

// Forced alignment (WhisperX-style) — Turkish audio + transcript
const logits = await model.session.executor.extractLogits(turkishAudio);
const aligner = createWav2Vec2AlignerFromLogits(logits);
const alignment = aligner.align({ transcript: 'türkçe transcript metni' });

Files

• wav2vec2-large-xlsr-turkish.onnx — ONNX model graph (3.4 MB)
• wav2vec2-large-xlsr-turkish.onnx.data — External weight data (1.2 GB)
• config.json — Original HuggingFace model config
• vocab.json — Turkish character vocabulary with CTC blank token

Source

Based on facebook/wav2vec2-large-xlsr-53, fine-tuned by m3hrdadfi on Common Voice Turkish.