Transcribe English audio locally with Granite Speech 5 TurboCTC

Granite Speech 5.0 TurboCTC is an automatic speech recognition¹ model in IBM's Granite Speech family. It takes audio and returns text; the 5.0 TurboCTC model card describes English speech recognition. For Korean or Japanese meetings, choose an ASR system that supports those languages. Official model card

Install PyTorch², an audio-input library, and Transformers in a virtual environment³. Copy an English WAV into the working folder as sample.wav, then save the code as transcribe.py and run it to print the transcript.

python -m pip install torch torchaudio soundfile "transformers>=5.16.0" datasets

import soundfile as sf
from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="ibm-granite/granite-speech-5.0-470m-turboctc")
audio, sample_rate = sf.read("sample.wav", dtype="float32")
if audio.ndim > 1:
    audio = audio.mean(axis=1)
result = asr({"array": audio, "sampling_rate": sample_rate})
print(result["text"])

python transcribe.py

Listen to a short sample and compare sentences, numbers, and names with the output, then test a file at your real working length. Read IBM's measurement hardware and audio conditions alongside any reported comparison. Do not treat H200 or L4 results as desktop-GPU or laptop timings.

The model card lists an Apache 2.0 license. Check the terms for the model version and intended use, including redistribution if you publish results or bundle the model in a product. This guide treats it as English-only; for Korean, compare models that explicitly list Korean, such as Nemotron or Whisper variants.

If you need to transcribe Korean or Japanese audio, do not choose this model. It targets English speech recognition, so compare other languages with ASR models that name them.