OpenASR/whisper-medium

Whisper Medium · OpenASR

High-accuracy multilingual Whisper at 769M parameters

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

---

✨ Highlights

• 🎧 Multilingual ASR — transcribes many languages and can translate speech to English
• 🎯 769M parameters — near-large accuracy with a more manageable footprint
• 🌐 Weak-supervision scale — trained with Whisper's 680k-hour labelled speech corpus
• 🦀 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-medium:q8

# 3. Transcribe
openasr transcribe audio.wav --model whisper-medium

All builds for this model:

openasr pull whisper-medium:fp16
openasr pull whisper-medium:q8
openasr pull whisper-medium:q4

📦 Available builds

Quant	File (.oasr)	Size	RAM peak	RTF · M1 CPU	RTF · M1 GPU	JFK ΔWER vs fp16
fp16	whisper-medium-fp16.oasr	1.53 GB	4.03 GB	0.62×	0.61×	0.0%
q8_0	whisper-medium-q8_0.oasr	874 MB	2.17 GB	0.46×	0.41×	0.0%
q4_k	whisper-medium-q4_k.oasr	522 MB	1.54 GB	0.51×	0.39×	0.0%

_{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 Medium

Whisper Medium is OpenAI's 769M-parameter multilingual Whisper checkpoint. It uses the standard Whisper encoder-decoder architecture for automatic speech recognition and speech translation, trained with large-scale weak supervision on 680k hours of labelled speech. Medium delivers much of the large model's accuracy at a smaller footprint, a strong choice when quality matters but the largest checkpoint is too heavy. This OpenASR repo repackages the original openai/whisper-medium 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-medium with the OpenASR importer:

openasr model-pack import-whisper-local <src> <out>.oasr \
  --package-id whisper-medium --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: Apache-2.0 (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 Medium, released by OpenAI (openai/whisper-medium). All credit for the original model, training recipe, and weights belongs to OpenAI. The upstream Hugging Face model card declares Apache-2.0 licensing; 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-medium