Whisper Large V2

Whisper large-v2 is an automatic speech recognition (ASR) and speech translation model based on a Transformer encoder-decoder architecture, trained on 680,000 hours of weakly supervised multilingual data. It is a 1.55 billion parameter model that performs speech recognition, speech translation, and language identification without requiring fine-tuning for new domains. The model uses a sequence-to-sequence approach where tasks are specified via context tokens, enabling a single model to handle multilingual transcription, translation between languages, and timestamp prediction. It processes audio segments up to 30 seconds natively and supports arbitrary-length audio through chunking.

Key Strengths

Whisper large-v2 is trained for 2.5x more epochs than the original large model with added regularization, improving robustness. It generalizes to unseen datasets in a zero-shot manner, matching or exceeding prior fully supervised results on standard benchmarks without fine-tuning.

Benchmark Performance

On the Open ASR Leaderboard, large-v2 achieves a mean Word Error Rate (WER) of 7.835 across multiple datasets, with a Real-Time Factor (RTFX) of 144.452. Specific dataset results include:

Dataset	WER
LibriSpeech clean	2.83
Earnings22	12.05
Gigaspeech	10.67
AMI	16.74

On LibriSpeech test-clean, the model achieves a WER of approximately 3.0% in a zero-shot setting.

Architecture and Capabilities

The model supports multilingual speech recognition across 99 languages, speech translation from any supported language into English, and language identification. It processes audio natively in 30-second segments and supports arbitrary-length transcription through chunking with batched inference and optional timestamp prediction.

Performance Characteristics

Large-v2 requires approximately 10 GB of VRAM for inference. Its relative inference speed is 1x (baseline), compared to smaller variants which offer faster throughput at reduced accuracy. The model is available through gigarouter as a hosted, OpenAI-compatible API, eliminating the need for local GPU infrastructure or model management.