zeromodels/granite-speech-5.0-470m-turboctc

See our collection for all versions of Granite Speech.

Run Granite Speech 5.0 with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/granite-speech-5.0-470m-turboctc

Paper: Granite-speech: open-source speech-aware LLMs with strong English ASR capabilities (arXiv:2505.08699) · HF Papers

Granite Speech 5.0 turboctc is a fast, non-autoregressive CTC ASR model: a conformer encoder (block-wise self-attention with Shaw relative positions, two early time-subsampling blocks) with a self-conditioned CTC head. The whole utterance is transcribed in a single forward pass — no decoder loop — then greedy-CTC decoded. Output is lowercase subword text.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of ibm-granite/granite-speech-5.0-470m-turboctc for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an ASR (CTC) checkpoint (GraniteSpeech5CTC, 470M).

✨ Quick start

import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

import soundfile as sf
from zeromodels.models.granite_speech5 import (
    GraniteSpeech5CTC,
    GraniteSpeech5FeatureExtractor,
    GraniteSpeech5Tokenizer,
)

model = GraniteSpeech5CTC.from_weights("zeromodels/granite-speech-5.0-470m-turboctc")
features = GraniteSpeech5FeatureExtractor()
tokenizer = GraniteSpeech5Tokenizer.from_weights("zeromodels/granite-speech-5.0-470m-turboctc")

audio, sr = sf.read("your_audio.wav", dtype="float32")  # 16 kHz mono
inputs = features(audio, sampling_rate=sr)
predicted_ids = model.generate(inputs)          # greedy CTC decoding
print(tokenizer.batch_decode(predicted_ids))

Load any Granite Speech 5.0 variant the same way with from_weights("zeromodels/<variant>"):

Variant	Hub	Notes
granite-speech-5.0-470m-turboctc	zeromodels/granite-speech-5.0-470m-turboctc	470M · CTC ASR

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• 16 kHz mono audio; the feature extractor returns input_features + attention_mask.
• CTC decode collapses repeats and drops the blank token — tokenizer.batch_decode(...) handles it.
• See Granite Speech docs and Loading Weights.
• Community / upstream safetensors still work via the hf: prefix, e.g. GraniteSpeech5CTC.from_weights("hf:ibm-granite/granite-speech-5.0-470m-turboctc").

Special Thanks

A huge thank you to the IBM Granite team for creating and releasing these models.

License: Apache 2.0.