tujia-xlsr53-ft

Fine-tuned XLSR-53 encoder for Tujia (Northern dialect). This is only the encoder; it is not a complete ASR model and produces no transcript on its own. It is the feature extractor required by the tujia-ctc v0.2 syllable recognizer.

• Base model: facebook/wav2vec2-xlsr-53-espeak-cv-ft (Apache-2.0)
• Fine-tuning: the top 4 of 12 transformer layers were trainable (freeze_layers = 8) on ~2.6 hours of Tujia read speech, jointly with a 709-class toned-syllable CTC head
• Size: 315,438,720 parameters, fp32, ~1.26 GB across 5 standard safetensors shards
• Output: 1024-dim frame features at 50 Hz, identical interface to the base model

Do not substitute the original XLSR-53

The features of this encoder differ from facebook/wav2vec2-xlsr-53-espeak-cv-ft. Using the original weights with the tujia-ctc v0.2 head degrades recognition substantially; the head was trained against these features.

Usage

from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2Model

feat = Wav2Vec2FeatureExtractor.from_pretrained("xbaibing/tujia-xlsr53-ft")
enc = Wav2Vec2Model.from_pretrained("xbaibing/tujia-xlsr53-ft")

The v0.2 release archive fetches this repository through fetch_encoder.sh, which places it in models/tujia-xlsr53-ft and points TUJIA_XLSR at it.

Data and license

• Upstream weights: Apache-2.0, attribution must be retained.
• These fine-tuned weights: research use only. No rights to the training recordings or transcripts are granted, and neither is distributed.
• Parent project: https://github.com/ibaibing/TujiaNLP
• Model page: https://github.com/ibaibing/tujia-asr