wav2vec2-large-xlsr-korean

Converts Korean audio waveforms to text using a wav2vec2 architecture pretrained on 53 languages via XLSR (Cross-Lingual Speech Representations) and fine-tuned on the Zeroth Korean dataset. The model uses self-supervised learning on raw audio to learn acoustic representations, then applies a language-specific linear projection layer trained on Korean speech data to map acoustic features to Korean phonemes and words. Processes raw PCM audio at 16kHz sample rate through a convolutional feature extractor followed by transformer encoder blocks.

Extracts learned acoustic representations from raw audio using the wav2vec2 encoder backbone without the final classification head. The model applies a convolutional feature extractor (7 layers, 512 channels) to downsample raw waveforms, then passes through 12 transformer encoder layers with attention mechanisms to produce contextualized acoustic embeddings. These embeddings capture phonetic and speaker information in a 768-dimensional space, useful for downstream tasks beyond transcription.

Enables adaptation of the pretrained wav2vec2 model to domain-specific Korean speech by unfreezing the classification head and optionally the encoder layers, then training on custom labeled audio data. The model uses CTC (Connectionist Temporal Classification) loss to align variable-length audio sequences with Korean text transcriptions without requiring forced alignment. Supports mixed-precision training and gradient accumulation for efficient training on consumer GPUs.

Processes multiple Korean audio samples of different lengths in a single batch using dynamic padding and attention masks. The model pads shorter sequences to match the longest sequence in the batch, applies attention masks to ignore padding tokens, and processes all samples through the encoder in parallel. This approach maximizes GPU utilization and reduces per-sample inference latency compared to processing audio sequentially.

Enables real-time Korean speech-to-text transcription by processing audio in fixed-size chunks (e.g., 1-2 second windows) with overlap to maintain context. The model maintains a sliding buffer of recent audio frames, processes new incoming chunks through the encoder, and outputs partial transcriptions incrementally. Requires careful management of attention context across chunk boundaries to avoid artifacts at segment boundaries.

Leverages cross-lingual speech representations learned from 53 languages during XLSR pretraining to improve Korean ASR performance with limited labeled data. The model's encoder has learned language-agnostic acoustic patterns (phoneme-like units, prosody, speaker characteristics) that transfer effectively to Korean. Fine-tuning only the task-specific CTC head requires minimal Korean data compared to training from scratch.