whisper-large-v3-turbo

<div align="center">

Whisper Large v3 Turbo · OpenASR

Fast multilingual Whisper built from pruned large-v3

Native speech-to-text in the OpenASR runtime — engineered for peak performance on CPU & GPU, no Python at inference time.

</div>

---

✨ Highlights

• ⚡ Turbo decoder — prunes Whisper large-v3's decoder from 32 layers to 4 for much faster generation
• 🌍 Multilingual ASR — transcribes many languages and can translate speech to English
• 🎙️ Zero-shot robustness — inherits Whisper's large-scale weak-supervision training across noisy domains
• 🦀 Native in OpenASR — .oasr packs run with no Python at inference, engineered for peak performance on CPU & GPU

🚀 Quickstart

# 1. Install the OpenASR CLI  ·  https://openasr.org
# 2. Pull a build (pick a quant — see the table below)
openasr pull whisper-large-v3-turbo:q8

# 3. Transcribe
openasr transcribe audio.wav --model whisper-large-v3-turbo

All builds for this model:

openasr pull whisper-large-v3-turbo:fp16
openasr pull whisper-large-v3-turbo:q8
openasr pull whisper-large-v3-turbo:q4

📦 Available builds

_{RTF = real-time factor on the fixed 11s JFK clip (lower is faster); RAM peak measured per pack in an isolated subprocess. JFK ΔWER compares each quantized build's JFK transcript to this model's fp16 JFK transcript, so it measures quantization drift rather than absolute recognition accuracy. q8_0 is the recommended default — near-reference quality at a fraction of the footprint.}

🧠 About Whisper Large v3 Turbo

Whisper Large v3 Turbo is OpenAI's faster variant of Whisper large-v3: it keeps the same Whisper architecture and multilingual speech-recognition/translation interface, but reduces the decoder depth from 32 layers to 4. The upstream card describes the result as much faster with only a minor quality trade-off, while retaining Whisper's broad zero-shot behavior from training on more than five million hours of labeled audio. This OpenASR repo repackages the original openai/whisper-large-v3-turbo weights as .oasr packs that run natively in the OpenASR runtime with no Python at inference time. For most users the q8_0 build is the recommended default; q4_k is for tighter memory budgets and fp16 is for verification or maximum fidelity.

⚙️ How these packs were made

Converted from openai/whisper-large-v3-turbo with the OpenASR importer:

openasr model-pack import whisper <src> <out>.oasr \
  --package-id whisper-large-v3-turbo --quantization {fp16,q8-0,q4-k}

The .oasr container is GGUF-backed; packs use zero-copy mmap weight binding and graph buffer reuse to keep peak memory low.

⚖️ License

These packs inherit the upstream model's license: MIT (source). OpenASR packaging retains the upstream copyright and NOTICE; the only modifications are format conversion and quantization.

🙏 Acknowledgements

This pack is a redistribution of Whisper Large v3 Turbo, released by OpenAI (openai/whisper-large-v3-turbo). All credit for the original model, training recipe, and weights belongs to OpenAI. The packs inherit the upstream MIT license; OpenASR only converts the weights into .oasr packages and adds quantized builds for local runtime use.

🔗 Links

• 🦀 OpenASR — https://github.com/QuintinShaw/openasr
• 🌐 Website — https://openasr.org
• 🤗 Upstream model — openai/whisper-large-v3-turbo