Granite Speech 5.0 470M TurboCTC Multilingual GGUF
GGUF conversions of ibm-granite/granite-speech-5.0-470m-turboctc for use with transcribe.cpp.
This conversion uses the updated multilingual release from October 21, 2026, not the earlier English-only checkpoint.
Supported languages: English, French, German, Spanish, Portuguese, Italian, and Japanese.
Download
| File | Quantization | Size |
|---|---|---|
| granite-speech-5.0-470m-turboctc-multilingual-Q8_0.gguf | Q8_0 | 505,663,136 bytes |
SHA-256:
f3efdc05b85b3a764bb5bb7022e69f61cafcecc11605c881c6396a6ed07ec91b
Usage
Use a transcribe.cpp build with Granite Speech 5.0 TurboCTC support. For example:
build/bin/transcribe-cli \
-m granite-speech-5.0-470m-turboctc-multilingual-Q8_0.gguf \
audio.wav
The model is for speech recognition, not translation. Its input sample rate is 16 kHz. Multilingual recognition does not mean that the runtime returns a detected language: stt.capability.lang_detect remains false.
Conversion Provenance
- Source checkpoint: ibm-granite/granite-speech-5.0-470m-turboctc at 1dd23f46f6d4e6eae2531982ba2e33220086dd97.
- Converter/runtime revision: transcribe.cpp at 4d7c1ccefd39e5a9ea36cca5f8fefc6dcc14b99b.
- Conversion: source weights to diagnostic F32 GGUF, then
transcribe-quantize --quant Q8_0. - The converter patch records all seven supported languages and accepts the updated processor metadata without requiring the optional legacy
preprocessor_config.json. - The original multilingual vocabulary and
<unk>CTC blank token are preserved. This is a format conversion and quantization, not a fine-tune.
See conversion.json and the converter patch for the exact provenance.
Verification
On three fixed clips, one each in English, Spanish and Japanese, this Q8_0 file produced exactly the same transcript as a fresh Transformers F32 reference using identical mono 16 kHz PCM input.
This is a limited conversion-parity check, not a comprehensive multilingual WER benchmark. The other supported languages were not independently evaluated in that check.
License And Attribution
The original model weights are provided by IBM Granite under Apache 2.0. Conversion and quantization packaging: 17slever17.
The converter patch derives from the MIT-licensed transcribe.cpp project; its license is preserved separately in conversion/LICENSE-transcribe.cpp. This repository is not an official IBM release.