OpenASR/granite-speech-4.1-2b

Granite Speech 4.1 2B · OpenASR

IBM Granite Speech 4.1 — a 2B multilingual ASR model with strong English accuracy

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

---

✨ Highlights

• 📊 Verified English WER: 0.54% — identical across fp16 / q8_0 / q4_k on OpenASR's frozen evaluation set
• 🌐 Six trained languages — English, French, German, Spanish, Portuguese, Japanese (auto-detected; no language token)
• 📦 Three quantization tiers: fp16 / q8_0 / q4_k — delivered in OpenASR's native .oasr format for local inference
• 🦀 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 granite-speech-4.1-2b:q8

# 3. Transcribe
openasr transcribe audio.wav --model granite-speech-4.1-2b

All builds for this model:

openasr pull granite-speech-4.1-2b:fp16
openasr pull granite-speech-4.1-2b:q8
openasr pull granite-speech-4.1-2b:q4

📦 Available builds

Quant	File (.oasr)	Size	RAM peak	RTF · M1 CPU	RTF · M1 GPU	JFK ΔWER vs fp16
fp16	granite-speech-4.1-2b-fp16.oasr	4.63 GB	5.91 GB	0.40×	0.23×	0.0%
q8_0	granite-speech-4.1-2b-q8_0.oasr	2.47 GB	3.65 GB	0.35×	0.17×	0.0%
q4_k	granite-speech-4.1-2b-q4_k.oasr	1.55 GB	2.63 GB	0.29×	0.15×	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 Granite Speech 4.1 2B

Granite Speech 4.1 2B is a speech-recognition model from IBM Granite. Its architecture pairs a 16-layer Conformer audio encoder with a Q-Former projector feeding a 2B Granite decoder. The model auto-detects the source language through free-text instruction-following and ships trained coverage for English, French, German, Spanish, Portuguese, and Japanese. OpenASR distributes this model in three quantization tiers -- fp16, q8_0, and q4_k -- packaged in the native .oasr runtime format for local inference.

⚙️ How these packs were made

Converted from ibm-granite/granite-speech-4.1-2b with the OpenASR importer:

openasr model-pack core: convert_local_granite_speech_source_to_runtime_pack <src> <out>.oasr \
  --package-id granite-speech-4.1-2b --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 granite-speech-4.1-2b, created and released by IBM Granite (ibm-granite/granite-speech-4.1-2b) under the Apache License 2.0. OpenASR performs format conversion, quantization, runtime validation, and local-inference adaptation only; all model weights and training are the work of the original authors.

🔗 Links

• 🦀 OpenASR — https://github.com/QuintinShaw/openasr
• 🌐 Website — https://openasr.org
• 🤗 Upstream model — ibm-granite/granite-speech-4.1-2b