wav2vec2-large-xlsr-53-portuguese

Converts Portuguese audio (16kHz mono WAV format) to text using wav2vec2 architecture with XLSR-53 cross-lingual pretraining. The model uses a self-supervised learning approach where it first learns universal speech representations from 53 languages via masked prediction on unlabeled audio, then fine-tunes on Portuguese Common Voice 6.0 dataset (validated splits only). Inference runs via HuggingFace Transformers pipeline or direct model loading, accepting raw audio tensors and outputting character-level transcriptions with optional confidence scores.

Processes multiple Portuguese audio files sequentially or in mini-batches through the wav2vec2 pipeline, automatically handling audio resampling (to 16kHz), normalization, and padding. Implements error recovery for corrupted files, mismatched sample rates, and out-of-memory conditions. Returns structured output mapping input file paths to transcriptions with per-file processing status and optional timing metrics.

Enables further fine-tuning of the pretrained wav2vec2-xlsr-53 checkpoint on custom Portuguese audio datasets using the HuggingFace Trainer API. Implements CTC loss (Connectionist Temporal Classification) for sequence-to-sequence alignment, with support for mixed-precision training (fp16) and gradient accumulation for memory efficiency. Includes data collation for variable-length audio, automatic vocabulary building from transcripts, and evaluation metrics (WER, CER) on validation splits.

Extracts learned audio representations (embeddings) from intermediate layers of the wav2vec2 model, enabling use as features for downstream tasks beyond transcription. The model outputs 768-dimensional embeddings per audio frame (at 50Hz temporal resolution) from the transformer encoder, which can be pooled or aggregated for speaker identification, emotion detection, language identification, or audio classification. Representations are frozen (no gradient flow) unless explicitly fine-tuned.

Implements streaming speech recognition by processing audio in fixed-size chunks (e.g., 1-second windows) and maintaining a sliding buffer of context frames for the transformer encoder. Each chunk is independently transcribed with optional context from previous frames to improve accuracy on chunk boundaries. Outputs partial transcriptions incrementally as audio arrives, with final transcription refinement when audio stream ends.

Converts the full-precision (fp32) wav2vec2 model to reduced-precision formats (int8, fp16, or dynamic quantization) for deployment on resource-constrained devices (mobile, embedded systems, edge servers). Quantization reduces model size by 4-8x and inference latency by 2-3x with minimal accuracy loss (<1% WER increase). Supports ONNX export for cross-platform deployment and TensorRT optimization for NVIDIA hardware.