Granite Speech 5.0 TurboCTC 470M GGUF

Granite Speech 5.0 TurboCTC 470M GGUF

Converts speech to text in seven languages, as quantized GGUF versions of IBM's Granite Speech 5.0 TurboCTC 470M, built for transcribe.cpp. For developers adding local dictation or transcription.

Made by vladers on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Granite Speech 5.0 TurboCTC 470M GGUF is a set of speech-to-text files made by vladers from IBM's Granite Speech model of the same name. It transcribes seven languages, English among them, and writes text with punctuation and capital letters. The files target the transcribe.cpp program and come in seven versions, from 4-bit to full 32-bit precision.

Who it suits: It suits developers who add dictation or transcription to an app that runs on the user's own computer, and people who want a compact recognizer whose build can be repeated and checked.

Industry | Who uses it | How they use it | Example
Software | App developers | Add offline dictation to a desktop app | Use the Q8_0 file the way Voiceour does
Media | Editors and podcasters | Turn recorded interviews into text | Transcribe a 16 kHz WAV file
Research | Speech researchers | Compare file sizes by word error rate | Read the per-file results in the card

Anyone who needs languages beyond those seven, or a finished app instead of model files, can skip it.

Where to use it:

• transcribe.cpp, a C++ speech program that you build with CMake
• A Windows, macOS or Linux computer with Python for the download
• Apps that embed it, such as the macOS dictation app Voiceour

Use case | Who | What they get back
Local dictation | App developers | Text from speech, with punctuation
Interview transcripts | Editors | Written text from audio recordings
Quantization checks | Researchers | Error rates for each file size

How to set it up:

• Python with pip, for the download
• A C++ build setup with CMake, plus ffmpeg to convert audio
• Disk space for the file you pick, from about 279 MB for Q4_K_M to 1.89 GB for F32
• The transcribe.cpp source code, taken from its GitHub page

1. Install the Hugging Face hub client.
2. Download one file, such as the Q8_0 one.
3. Get the transcribe.cpp source code into a folder named transcribe.cpp, then build its command line tool with CMake.
4. Run the tool on a 16 kHz mono WAV file.

pip install -U huggingface_hub
hf download vladers/granite-speech-5.0-470m-turboctc-GGUF granite-speech-5.0-470m-turboctc-Q8_0.gguf --local-dir .

cmake -S transcribe.cpp -B build
cmake --build build --target transcribe-cli -j

How to use it: Convert your recording to a 16 kHz mono WAV with ffmpeg, then give the tool the model file and the audio. The card advises reading its Known behaviour notes before you try silence or tiny clips.

ffmpeg -i in.mp3 -ar 16000 -ac 1 audio.wav

build/bin/transcribe-cli -m granite-speech-5.0-470m-turboctc-Q8_0.gguf audio.wav

Tool | What it does
ffmpeg | Converts an audio file to 16 kHz mono WAV
transcribe-cli | Runs the model file on the audio and returns text
hf download | Fetches the model file from Hugging Face

Example: You build transcribe-cli, download the Q8_0 file, convert an MP3 interview with ffmpeg and run the tool on it. The card does not show a sample printout, so check the tool's own help for the output format.

For newcomers and experts: Speech recognition turns audio into text, and a smaller file trades a little accuracy for less disk and memory. Start with the Q8_0 file, which the card calls the default choice.

Of the 520 tensors in each file, 163 are compressed, namely the linear layers and pointwise convolutions, while the other 357 stay at full 32-bit precision. The BF16 file keeps IBM's original linear weights. The card measures every file on 19,330 English clips and lists speed on an Apple M4 Pro.

IBM's Granite Speech checkpoint
Converted to seven GGUF files
Pick a file by size and accuracy
transcribe.cpp loads it
Audio goes in and text comes out

Title: Word error rate on the confirmation set in percent
F32: 5.30%
Q8_0: 5.30%
Q6_K: 5.29%
Q5_K_M: 5.32%
Q4_K_M: 5.39%
Parakeet TDT 0.6B v3: 7.21%

Title: File size by version in MB
Q4_K_M: 279 MB
Q5_K_M: 336 MB
Q6_K: 392 MB
Q8_0: 506 MB
F16: 948 MB

Good to know:

• Versions: seven files, from Q4_K_M at 279 MB to F32 at 1.89 GB
• Accuracy: on the 17,858-clip English confirmation set, Q4_K_M scores 5.39 percent word error against 5.30 for the full-precision file
• Size: the record's rough 4-bit estimate is about 237 MB, while the card lists 279 MB for Q4_K_M
• Languages: seven, and the accuracy tests cover English only
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the audio you process on your own computer, while the pip and git steps fetch packages and code from the web
• Not for: languages outside the seven, or a finished app with a window

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Parameters: 473M
• Weights: about 237 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 431
• Likes: 4
• License: apache-2.0
• Languages: en, fr, de, es, pt, it, ja
• Based on: ibm-granite/granite-speech-5.0-470m-turboctc
• Tags: gguf, transcribe.cpp, speech-recognition, granite, automatic-speech-recognition, multilingual

Links

• Model: https://huggingface.co/vladers/granite-speech-5.0-470m-turboctc-GGUF
• Files: https://huggingface.co/vladers/granite-speech-5.0-470m-turboctc-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download vladers/granite-speech-5.0-470m-turboctc-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.