Releases·Model Release
IBM details Granite 5.0 TurboCTC: 470M-parameter encoder-only ASR model on Open ASR speed-accuracy Pareto frontier
IBM researchers describe Granite 5.0 TurboCTC, a 470 million-parameter encoder-only CTC speech recognition model. It uses a Conformer-based architecture with training and inference optimizations to reach the speed-accuracy Pareto frontier of the Open ASR leaderboard for English short-form ASR, running twice as fast as the fastest competitor. The model is trained on only publicly available data and released under a permissive license on Hugging Face.
0 engagement1 sourceThu, Sep 17, 2026, 12:05 PM
AI Overview
- IBM's Granite 5.0 TurboCTC is a 470M-parameter encoder-only Conformer model using pyramidal subsampling, chunk-wise attention, and mid-layer prediction conditioning
- Training uses only publicly available data, a novel Muon optimizer, and balanced data sampling
- Inference speedups come from replacing 1x1 convolutions with linear layers and optimizing Conformer attention
- The model sits on the Open ASR leaderboard's English short-form Pareto frontier while running twice as fast as the fastest competitor
- The model ships under a permissive license on Hugging Face as ibm-granite/granite-speech-5.0-470m-turboctc
Related
Releases·Model Release
IBM releases Granite Speech 5.0 Turbo CTC for fast transcription
IBM released Granite Speech 5.0 Turbo CTC, a speech recognition model optimized for speed and accuracy. The model is available on Hugging Face, offering practitioners a new option for efficient transcription.
IBM released Granite Speech 5.0 Turbo CTC, a speech recognition model optimized for speed and accuracy. The model is available on Hugging Face, offering practitioners a new option for efficient transcription.
154 engagement1 sourceredditTue, Aug 25, 2026, 07:44 PM
Releases·Model Release
Microsoft releases VibeVoice-ASR-BitNet, a compressed ASR model for edge CPU inference
Microsoft released VibeVoice-ASR-BitNet, a compressed variant of VibeVoice-ASR optimized for real-time inference on edge CPUs without a GPU. The model uses heterogeneous quantization to shrink from 4.62 GB to 1.58 GB, achieving 1.6–2.3× faster inference than Whisper.cpp with real-time capability on as few as 3 CPU threads.
Microsoft released VibeVoice-ASR-BitNet, a compressed variant of VibeVoice-ASR optimized for real-time inference on edge CPUs without a GPU. The model uses heterogeneous quantization to shrink from 4.62 GB to 1.58 GB, achieving 1.6–2.3× faster inference than Whisper.cpp with real-time capability on as few as 3 CPU threads.
43 engagement1 sourceredditTue, Jul 28, 2026, 04:10 AM
Releases·Model Release
Orukeet ASR model, built from NVIDIA Parakeet TDT 0.6B v3, outperforms Parakeet on 61 of 74 test splits
Orukeet, a 25-language speech recognizer derived from NVIDIA Parakeet TDT 0.6B v3, is available on Hugging Face. It replaces half the encoder's temporal depthwise filters with 12,288 fitted, frozen Gabor kernels and outperforms Parakeet on 61 of 74 tested splits, with particularly strong gains on multilingual and multi-accent data.
Orukeet, a 25-language speech recognizer derived from NVIDIA Parakeet TDT 0.6B v3, is available on Hugging Face. It replaces half the encoder's temporal depthwise filters with 12,288 fitted, frozen Gabor kernels and outperforms Parakeet on 61 of 74 tested splits, with particularly strong gains on multilingual and multi-accent data.
108 engagement1 sourceredditFri, Sep 11, 2026, 03:50 PM
Developments·Hardware
I-Parakeet: Integer-Only Conformer ASR on Mobile NPU
Researchers propose I-Parakeet, an integer-only implementation of NVIDIA's Parakeet-CTC (0.6B parameters) that runs entirely on a smartphone NPU without floating-point operators or CPU fallback. The work targets the common problem of quantized Conformer ASR models falling back to floating point on numerically sensitive operations. It matters for practitioners deploying ASR on edge devices, as it fully exploits integer accelerators like mobile NPUs.
Researchers propose I-Parakeet, an integer-only implementation of NVIDIA's Parakeet-CTC (0.6B parameters) that runs entirely on a smartphone NPU without floating-point operators or CPU fallback. The work targets the common problem of quantized Conformer ASR models falling back to floating point on numerically sensitive operations. It matters for practitioners deploying ASR on edge devices, as it fully exploits integer accelerators like mobile NPUs.
0 engagement1 sourcearxivFri, Sep 25, 2026, 05:47 AM
Releases·Tool Release
llama.cpp adds support for IBM Granite SWA and MoE models via PR #25505
A pull request by gabe-l-hart adds support for IBM's Granite models with sliding window attention (SWA) and mixture-of-experts (MoE) architectures to llama.cpp. This enables local inference of these new Granite variants.
A pull request by gabe-l-hart adds support for IBM's Granite models with sliding window attention (SWA) and mixture-of-experts (MoE) architectures to llama.cpp. This enables local inference of these new Granite variants.
3 engagement1 sourceredditWed, Aug 19, 2026, 10:13 PM