OpenAI Releases Whisper Large-v3, Cutting Speech Recognition Errors 10-20% Across Languages

OpenAI has released Whisper large-v3, an update to its open-weight automatic speech recognition (ASR) and speech translation model, available now on Hugging Face. The model reduces word error rates by 10% to 20% compared to Whisper large-v2 across a wide range of languages, according to OpenAI.

What changed

Whisper large-v3 keeps the same underlying architecture (approximately 1.55 billion parameters) as large and large-v2, with two technical modifications:

• Input spectrogram resolution: increased from 80 to 128 Mel frequency bins, giving the model more granular frequency information to work with.
• New language token: Cantonese is now supported as a distinct language, separate from Mandarin.

Training data

The model was trained on 1 million hours of weakly labeled audio combined with 4 million hours of pseudo-labeled audio — the latter generated using Whisper large-v2 itself — for a combined dataset of 5 million hours. Training ran for 2.0 epochs over this mixture. This is consistent with OpenAI's original Whisper approach of large-scale weak supervision rather than fully hand-labeled datasets, first described in the paper "Robust Speech Recognition via Large-Scale Weak Supervision" by Alec Radford et al.

Capabilities and access

Whisper large-v3 performs both transcription (source and target language matching) and translation (target text in English regardless of source language), with automatic language detection. It supports sentence-level and word-level timestamps and handles audio of arbitrary length through either sequential or chunked long-form decoding strategies. The model has a native receptive field of 30 seconds per pass.

The model is fully open-weight and distributed through Hugging Face Transformers, installable via pip and usable through the standard pipeline API or direct model/processor calls. As an open model, there is no per-token API pricing — cost is determined by self-hosted compute (GPU inference with torch.float16 recommended) or third-party hosting providers.

What this means

Whisper large-v3 is not a frontier LLM release, but it is one of the more widely deployed open models in production ASR pipelines — used in transcription tools, captioning services, and voice-agent front ends across the industry. A 10-20% error reduction is meaningful at scale: for high-volume transcription workloads, that translates directly into fewer manual corrections and lower downstream error propagation into diarization, translation, or LLM-based summarization steps that consume Whisper's output. The addition of a dedicated Cantonese token also signals OpenAI's continued investment in language coverage for markets where prior Whisper versions conflated related but distinct languages. Because the model remains open-weight, its real-world impact will be shaped less by OpenAI's own roadmap and more by how quickly the open-source ecosystem (Hugging Face, faster-whisper, and inference providers like Groq and Together AI) integrates and optimizes it for latency and cost.