Downloads · 30 days
44
20% of all-time downloads
superflow-hq/granite-speech-5.0-470m-turboctc-nc-Q8_0-GGUF
granite-speech-5.0-470m-turboctc-nc-Q8_0-GGUF is a automatic speech recognition model from superflow-hq. Use it when you need speech turned into text. The card lists the license as cc-by-nc-sa-4.0.
A standalone Q80 GGUF conversion of IBM's ibm-granite/granite-speech-5.0-470m-turboctc-nc for native ASR inference with audio.cpp, a pure C++ audio inference runtime powered by ggml.
Downloads · 30 days
44
20% of all-time downloads
All-time downloads
223
Public
Repo size
506 MB
Likes
0
Public
Click a slice to open those files.
.gguf506 MB · 100%
From the Hugging Face model README
A standalone Q8_0 GGUF conversion of IBM's ibm-granite/granite-speech-5.0-470m-turboctc-nc for native ASR inference with audio.cpp, a pure C++ audio inference runtime powered by ggml.
This is an unofficial community conversion. It is not produced, endorsed, or maintained by IBM.
Non-commercial only. The source checkpoint is licensed under CC-BY-NC-SA-4.0, and this conversion keeps the same license. Do not use this model commercially. For commercial use, use IBM's Apache-2.0
granite-speech-5.0-470m-turboctccheckpoint instead.
| File | Purpose |
|---|---|
granite-speech-5.0-470m-turboctc-nc-q8_0.gguf | Standalone Q8_0 GGUF model |
convert_granite5_turboctc_to_gguf.py | Reproducible one-shot conversion script |
transcribe.py | Small Python wrapper around native audiocpp_cli ASR |
conversion-manifest.json | Sanitized conversion metadata, source revision, and hashes |
The GGUF was produced from IBM's BF16 Safetensors checkpoint using the current audio.cpp GGUF converter and the granite5asr model specification.
Conversion validation reported:
weight_type=q8_0
embedded_sidecars=true
embedded_model_spec=true
model_spec_family=granite5asr
tensors=550
rank0_scalars=16
Conversion details:
| Property | Value |
|---|---|
| Source model | ibm-granite/granite-speech-5.0-470m-turboctc-nc |
| Source revision | 0eb7b4fe726a294815dc45d342860465b5af68ef |
| Parameters | 470M |
| Source precision | BF16 |
| GGUF precision | Q8_0 |
| GGUF size | ~482.35 MiB |
| SHA-256 | fc7a5bfec66f8e38040e7f080f6d2a69c4cfd014961b1de6b66ec9c5cb93cd13 |
| Runtime family | granite5asr |
| Architecture | Conformer + CTC |
| Language | English |
IBM publishes two Granite Speech 5.0 TurboCTC checkpoints.
The NC checkpoint uses additional training data and IBM reports a slightly lower aggregate WER on the public OpenASR test sets.
| IBM Granite 5 checkpoint | License | Training audio | IBM-reported aggregate WER ↓ |
|---|---|---|---|
| 470M TurboCTC NC | CC-BY-NC-SA-4.0 | ~75k hours | 4.85% |
| 470M TurboCTC | Apache-2.0 | ~60k hours | 5.00% |
That is why this repository converts the NC variant.
Lower WER means fewer transcription errors, but one benchmark number does not tell the whole story. Microphone distance, room noise, speaker movement, accent, domain vocabulary, decoding implementation, and runtime can all change real-world accuracy.
WER means Word Error Rate. Lower is better.
Every number in the table below is a WER percentage measured under a different audio condition. There is no separate "WER" column because the cells themselves are the WER values.
| Model | Near Field | Lab Meas. | Lab Sim. | High SNR | Mov. High | Mid SNR | Mov. Mid | Low SNR | Mov. Low |
|---|---|---|---|---|---|---|---|---|---|
| Mega-ASR | 3.9 | 20.3 | 18.9 | 7.2 | 8.5 | 14.1 | 15.3 | 28.4 | 30.7 |
| Qwen3-ASR-1.7B | 3.8 | 20.0 | 18.6 | 6.7 | 8.2 | 13.9 | 15.9 | 29.3 | 31.9 |
| cohere-transcribe-03-2026 | 4.3 | 21.4 | 20.2 | 7.8 | 8.8 | 15.2 | 17.0 | 32.8 | 33.9 |
| granite-speech-4.1-2b-nar | 4.1 | 23.1 | 22.1 | 8.2 | 9.1 | 16.4 | 18.4 | 31.9 | 33.3 |
| granite-speech-5.0-470m-turboctc-nc | 4.3 | 23.1 | 22.1 | 9.1 | 10.2 | 16.7 | 18.7 | 32.9 | 33.4 |
| granite-speech-4.1-2b | 4.3 | 23.3 | 22.4 | 8.2 | 9.3 | 16.4 | 18.6 | 36.6 | 36.7 |
| Qwen3-ASR-0.6B | 4.4 | 25.2 | 24.1 | 9.6 | 11.3 | 18.2 | 21.2 | 35.7 | 38.4 |
| granite-speech-5.0-470m-turboctc | 4.8 | 26.2 | 24.9 | 10.5 | 11.5 | 19.3 | 21.5 | 35.7 | 37.2 |
| parakeet-tdt-1.1b | 4.2 | 27.0 | 24.6 | 9.2 | 10.4 | 18.4 | 21.1 | 37.6 | 40.2 |
| parakeet-tdt-0.6b-v2 | 4.2 | 26.7 | 25.4 | 10.2 | 10.9 | 19.1 | 21.4 | 37.1 | 38.7 |
The scenario names are easier to understand than they initially look:
| Scenario | Meaning |
|---|---|
| Near Field | Microphone close to the speaker. Clean, direct speech. |
| Lab Meas. | Audio measured in controlled real-world acoustic conditions. |
| Lab Sim. | Similar acoustic conditions generated through simulation. |
| High SNR | Speech is much louder than the background noise. |
| Mov. High | Speaker or microphone movement under relatively clean conditions. |
| Mid SNR | More noticeable background noise. |
| Mov. Mid | Movement combined with moderate noise. |
| Low SNR | Difficult audio where background noise competes strongly with speech. |
| Mov. Low | Movement combined with low-SNR, difficult audio. |
SNR means Signal-to-Noise Ratio. Higher SNR generally means cleaner speech. Lower SNR means more noise relative to the speaker.
The row relevant to this repository is:
| Model | Near Field | Lab Meas. | Lab Sim. | High SNR | Mov. High | Mid SNR | Mov. Mid | Low SNR | Mov. Low |
|---|---|---|---|---|---|---|---|---|---|
| granite-speech-5.0-470m-turboctc-nc | 4.3 | 23.1 | 22.1 | 9.1 | 10.2 | 16.7 | 18.7 | 32.9 | 33.4 |
This gives a more useful picture than one aggregate number. The model is strong in clean near-field speech, while WER rises as noise, distance, and movement make recognition harder.
These benchmark values describe the underlying model evaluation. They are not measurements of this Q8_0 GGUF conversion unless the exact quantized file is evaluated with the same benchmark harness.
The 4.85% aggregate WER is IBM's result for the original model, not a measured WER for this Q8_0 GGUF.
This repository does not claim that the Q8_0 conversion has exactly 4.85% WER.
Quantization changes numerical precision and can change model predictions. The conversion has been structurally validated and inspected, but a Q8-specific WER number should only be published after evaluating this exact GGUF with the same benchmark and evaluation pipeline.
In other words: the architecture, weights, metadata, model specification, and runtime compatibility were validated. Benchmark parity should not be assumed without measurement.
Granite Speech 5.0 TurboCTC is not an LLM. It is an English automatic speech recognition model built around a Conformer encoder and CTC decoding.
The architecture uses:
The source model was released by IBM on August 25, 2026.
This GGUF targets audio.cpp's native granite5asr implementation.
audio.cpp provides the inference graph needed to run Granite Speech 5's Conformer + CTC architecture through native C++ and ggml.
Do not advertise or load this as a llama.cpp model.
GGUF is a file format, not a universal runtime adapter. llama.cpp does not currently implement the GraniteSpeech5ForCTC Conformer + CTC inference graph.
Giving a model a .gguf extension does not magically make every GGUF runtime understand its architecture. Computers remain disappointingly literal about these things.
This model is intended for audio.cpp, which provides native support for the Granite 5 ASR architecture.
The easiest current route on Apple Silicon is audio.cpp with Metal.
brew tap 0xShug0/audio-cpp
brew trust 0xShug0/audio-cpp
brew install audio-cpp
Or build the current source:
git clone https://github.com/0xShug0/audio.cpp.git
cd audio.cpp
scripts/build_metal.sh --target audiocpp_cli
The built CLI is normally located at:
build/macos-metal-release/bin/audiocpp_cli
Using the Hugging Face CLI:
pip install -U huggingface_hub
hf download superflow-hq/granite-speech-5.0-470m-turboctc-nc-Q8_0-GGUF \
granite-speech-5.0-470m-turboctc-nc-q8_0.gguf \
--local-dir .
Repository:
superflow-hq/granite-speech-5.0-470m-turboctc-nc-Q8_0-GGUF
Metal:
audiocpp_cli \
--task asr \
--family granite5asr \
--model ./granite-speech-5.0-470m-turboctc-nc-q8_0.gguf \
--backend metal \
--audio ./sample.wav
CPU:
audiocpp_cli \
--task asr \
--family granite5asr \
--model ./granite-speech-5.0-470m-turboctc-nc-q8_0.gguf \
--backend cpu \
--audio ./sample.wav
Long-form and VAD behavior can be controlled through audio.cpp request options. Check the upstream Granite 5 audio.cpp documentation for the current options.
Python is optional.
The actual model inference still runs in native C++ through audiocpp_cli. The included Python script is only a small wrapper around that runtime.
python3 transcribe.py sample.wav \
--model ./granite-speech-5.0-470m-turboctc-nc-q8_0.gguf \
--backend metal
python3 transcribe.py sample.wav \
--repo superflow-hq/granite-speech-5.0-470m-turboctc-nc-Q8_0-GGUF \
--backend metal
If audiocpp_cli is not available on PATH:
python3 transcribe.py sample.wav \
--model ./granite-speech-5.0-470m-turboctc-nc-q8_0.gguf \
--cli /absolute/path/to/audiocpp_cli \
--backend metal
Put a Hugging Face token in .env:
HF_TOKEN=hf_your_token_here
Then run:
python3 convert_granite5_turboctc_to_gguf.py
The converter:
audio.cpp converter;GraniteSpeech5ForCTC;For BF16:
python3 convert_granite5_turboctc_to_gguf.py --quant bf16
Base model:
This repository contains a quantized format conversion and preserves the source model's non-commercial licensing restrictions.
For commercial use, use IBM's Apache-2.0 sibling checkpoint instead:
https://huggingface.co/ibm-granite/granite-speech-5.0-470m-turboctc
IBM Granite Speech 5.0 TurboCTC release:
https://huggingface.co/blog/ibm-granite/granite-speech-5-0-470m-turboctc
IBM Granite Speech 5.0 470M TurboCTC NC:
https://huggingface.co/ibm-granite/granite-speech-5.0-470m-turboctc-nc
audio.cpp:
https://github.com/0xShug0/audio.cpp
Granite 5 support in audio.cpp:
https://github.com/0xShug0/audio.cpp/blob/main/docs/community_models/granite5asr.md
Hugging Face Open ASR leaderboard:
https://huggingface.co/datasets/hf-audio/open-asr-leaderboard
FFASR leaderboard:
https://huggingface.co/spaces/treble-technologies/ffasr
This repository is an independent community conversion.
IBM owns and maintains the original Granite Speech model. audio.cpp owns and maintains its runtime implementation.
No affiliation with or endorsement by IBM or audio.cpp is implied.