whisper

Converts audio input (WAV, MP3, M4A, FLAC, OGG) into text transcriptions using a Transformer-based encoder-decoder architecture trained on 680,000 hours of multilingual audio data. The model automatically detects the source language without explicit specification, then transcribes across 99 languages using a unified tokenizer. Inference runs via ONNX or PyTorch backends, with the Gradio interface handling audio upload, streaming, and real-time processing on HuggingFace Spaces infrastructure.

Automatically handles diverse audio input formats (MP3, M4A, FLAC, OGG, WAV) by normalizing to a standard 16kHz mono PCM stream before feeding to the Whisper model. The Gradio interface abstracts format detection and conversion using librosa or ffmpeg backends, transparently converting compressed or multi-channel audio without user intervention. This preprocessing ensures consistent model input regardless of source format or encoding.

Identifies the spoken language in audio without explicit user specification by using a language classification head trained as part of the Whisper model. The encoder processes the audio spectrogram and outputs language probabilities across 99 supported languages; the model selects the highest-confidence language and uses language-specific tokens to guide transcription. This enables single-pass processing without requiring separate language detection preprocessing.

Provides a Gradio-based web UI hosted on HuggingFace Spaces enabling users to upload audio files, trigger transcription, and view results in a browser without local setup. The interface handles file upload, displays transcription progress, and streams results back to the client. Gradio abstracts HTTP request handling, file management, and GPU resource allocation, allowing stateless inference on shared Spaces infrastructure with automatic scaling and timeout management.

Enables programmatic transcription of multiple audio files by importing the Whisper Python library and calling the transcribe() function in a loop or parallel batch. The local implementation uses PyTorch or ONNX backends, loading the model once and reusing it across files to amortize startup overhead. Developers can control model size (tiny, base, small, medium, large), language override, and output format (JSON with timestamps, plain text, SRT subtitles).

Provides five pre-trained model variants (tiny, base, small, medium, large) with different parameter counts (39M to 1.5B) allowing developers to select based on accuracy requirements and computational constraints. Smaller models (tiny, base) run faster on CPU and mobile devices but sacrifice transcription accuracy; larger models (medium, large) achieve higher accuracy but require GPU and more memory. The model selection is exposed via the Python API (whisper.load_model('base')) and can be configured in the Spaces demo via environment variables.

Generates transcription output with precise timestamps for each word or segment, enabling synchronization with video, subtitle generation, or audio-text alignment. The model outputs segment-level timestamps (start/end times in seconds) which can be further refined to word-level granularity via post-processing. The JSON output format includes timing information, allowing developers to build interactive transcripts, searchable video players, or automated subtitle tracks.