whisper-large-v3-turbo

Converts audio waveforms to text across 99 languages using a transformer-based encoder-decoder architecture trained on 680K hours of multilingual audio data. The model uses mel-spectrogram feature extraction from raw audio, processes variable-length sequences through a 24-layer encoder, and generates text tokens via an autoregressive decoder with cross-attention. Supports both streaming and batch inference modes with automatic language detection when language is not specified.

Identifies the spoken language in audio without explicit specification by analyzing mel-spectrogram features through the encoder's initial layers, which learn language-specific acoustic patterns. The model's multilingual token vocabulary includes language tokens that are predicted during decoding, allowing the system to infer language from phonetic and prosodic characteristics. Detection happens as a byproduct of transcription without separate inference passes.

Handles audio inputs of arbitrary duration (from seconds to hours) by converting to mel-spectrograms with fixed 80-dimensional frequency bins, then applying dynamic padding to 3000 time-steps (~30 seconds) or chunking longer sequences. The encoder processes padded sequences through 24 transformer layers with positional embeddings, while the decoder generates tokens autoregressively with a maximum output length of 448 tokens. Attention masks automatically handle padded regions to prevent information leakage.

Achieves noise robustness through training on 680K hours of diverse real-world audio including background noise, music, speech overlap, and poor recording conditions. The mel-spectrogram frontend acts as a lossy compression that emphasizes speech-relevant frequencies while attenuating noise. The encoder's deep transformer layers learn to suppress noise patterns through multi-head attention, which can focus on speech-dominant frequency bands. No explicit noise reduction preprocessing is required.

The Turbo variant achieves 2-3x faster inference than full Whisper v3 through knowledge distillation, where a smaller student model learns to mimic the full model's output distributions. The architecture uses the same transformer encoder-decoder design but with reduced layer depth and hidden dimensions, maintaining the 99-language capability through shared multilingual embeddings. Inference is further optimized through operator fusion and quantization-friendly design that enables INT8 quantization without accuracy loss.

Generates transcription output with precise timing information by tracking the decoder's attention alignment to the encoder's mel-spectrogram time-steps. Each generated token is associated with a start and end timestamp (in seconds) corresponding to the audio segment it represents. The alignment is computed through attention weights without requiring separate forced-alignment models, enabling end-to-end timing extraction in a single inference pass.

Processes multiple audio files simultaneously through batched tensor operations, with dynamic padding that groups audio of similar lengths to minimize wasted computation. The encoder processes all batch items in parallel through 24 transformer layers, while the decoder generates tokens autoregressively with cross-attention to the batch-encoded representations. Attention masks ensure each batch item only attends to its own padded sequence, preventing cross-contamination.