granite-speech-5.0-470m-turboctc

granite-speech-5.0-470m-turboctc

Turns speech in seven languages into punctuated, capitalized text. A small 470M-parameter model from IBM Granite for fast transcription on servers, laptops and phones.

Made by ibm-granite 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 is a speech-to-text model from IBM Granite. It listens to audio in English, French, German, Spanish, Portuguese, Italian or Japanese and writes the words with punctuation and capital letters.

It is small, at about 470 million parameters, and built for speed. It decodes in one pass instead of word by word, so it suits low-latency and high-throughput transcription.

Who it suits: Developers and teams who need accurate transcripts in several languages and want a model small enough to run on their own machines. The card names enterprise transcription as the intended use.

Industry | Who uses it | How they use it | Example
Software | App developers | Add speech input to a product | A note-taking app that turns voice memos into text
Customer service | Support teams | Transcribe recorded calls in several languages | A French and German call archive turned into searchable text
Media | Editors and producers | Draft transcripts of interviews and clips | A Japanese interview turned into a first-draft transcript
Devices | Mobile and edge engineers | Run transcription on the device itself | A phone app that transcribes without sending audio away

You can skip it if you need languages outside the seven listed, or if you want a ready-made app rather than a model to build on.

Where to use it: It runs wherever Python and the transformers library run, and the card also gives a route for Apple Silicon.

• Windows, macOS and Linux with Python and transformers
• Apple Silicon M series chips through mlx-audio
• Laptops, smartphones and other edge devices, as the card describes
• Servers that transcribe many files at once

Use case | Who | What they get back
Voice memo transcription | App builders | Punctuated text from short recordings
Multilingual call transcripts | Support and operations teams | Text in the speaker's own language
Edge speech input | Mobile developers | Low-latency text without a large model

How to set it up: You need a computer with Python and a recent transformers release. The card says version 5.16.0 or later supports the model natively.

• Python and pip
• transformers 5.16.0 or later, plus the datasets library
• Enough disk for about 946 MB of weights in BF16

1. Install Python if you do not have it.
2. Install transformers and datasets with the command below.
3. Optionally download the model files ahead of time with the Hugging Face commands.

pip install transformers>=5.16.0 datasets

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

How to use it: Load the processor and the CTC model, pass in your audio samples, call generate and decode the result. The card's own script does this on five short sample clips.

from datasets import Audio, load_dataset
from transformers import AutoModelForCTC, AutoProcessor

model_id = "ibm-granite/granite-speech-5.0-470m-turboctc"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCTC.from_pretrained(model_id, device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
speech_samples = [el["array"] for el in ds["audio"][:5]]

# `device` computes the log-mel front-end on the model's accelerator, saving a host-to-device copy
inputs = processor(
    speech_samples, sampling_rate=processor.feature_extractor.sampling_rate, device=model.device
)
inputs.to(model.device, dtype=model.dtype)
outputs = model.generate(**inputs)
print(processor.batch_decode(outputs, skip_special_tokens=True))

Step | What it does
AutoProcessor | Prepares the audio so the model can read it
AutoModelForCTC | Loads the speech model, placed on your GPU if there is one
generate | Produces the token output for each clip
batch_decode | Turns the output into readable text

Example: You feed it five short English clips from the sample dataset in the card's script. It prints a list of five transcripts, each with punctuation and capital letters.

For newcomers and experts: If you are new to this, start with the script above. It downloads a tiny test dataset, runs five clips through the model and prints the text, so you see a result before you try your own audio.

If you know the field: the encoder is a conformer that combines block attention, self-conditioning and downsampling over time. Its output layer covers 16,384 multilingual BPE units. It was trained with CTC on about 153,000 hours of public multilingual audio and decodes greedily without autoregression. That trades some modeling flexibility for speed.

Audio clip in
Conformer encoder reads it
CTC output layer picks tokens
Greedy decode, one pass
Punctuated text out

Good to know: The English average word error rate on the Open ASR leaderboard tests improved from the earlier English-only release, as the chart shows. The card reports Japanese character error rates separately and says the 470M model beats the 2B Granite model on all three Japanese test sets.

Title: Average English word error rate, percent
This multilingual release: 5.51 percent
Earlier English-only release: 5.78 percent

• License: Apache 2.0
• Cost to run: none named, the weights are free to download and run
• Data it reaches: the audio you give it, and the model files and test dataset it downloads from Hugging Face
• Not for: any language beyond the seven it lists

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 473M
• Weights: about 946 MB in BF16; about 236 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 54,890
• Likes: 117
• License: apache-2.0
• Languages: en, fr, de, es, pt, it, ja
• Quantized versions: 16
• Fine-tunes: 7
• Tags: transformers, safetensors, granite_speech5_ctc, automatic-speech-recognition, multilingual

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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