ibm-granite/granite-speech-5.0-470m-turboctc-nc

Granite Speech 5.0 470M TurboCTC NC

> Disclaimer: This model is intended for research and non-commercial use-only. See Granite Speech 5.0 Turbo CTC if interested in commercial use.

Model Summary: Granite Speech 5.0 TurboCTC Non-commercial is a compact 470 million parameter English ASR model with very high inference speed that is released for research and noncommercial purposes only. We invite users to refer to granite-speech-5.0-470m-turboctc for other use cases.

The model consists of a conformer acoustic encoder with block self-attention, self-conditioning and temporal downsampling with an output layer corresponding to 16,384 BPE units. It was trained on approximately 75,000 hours of English audio from public corpora using Connectionist Temporal Classification (CTC) and inference is done non-autoregressively with greedy decoding.

Evaluations:

We evaluated granite-speech-5.0-470m-turboctc-nc on standard short-form English ASR benchmarks from the Open ASR leaderboard:

!granite-speech-5.0-470m-turboctc-nc Performance on the Open ASR leaderboard (official results as of August 25, 2026, public test sets only, RTFx measured on 1 H200):

!wer_rtfx_nc !wer_size_nc

Performance on noisy and reverberant speech from the FFASR leaderboard (official results as of August 25, 2026, RTFx measured on 1 L4 GPU)

!ffasr_wer !ffasr_rtfx

Release Date: August 25, 2026

License: CC-BY-NC-SA-4.0

Supported Languages: English

Intended Use: Use of this model is governed by the terms of the CC-BY-NC-SA-4.0 license.

Usage:

Granite Speech 5.0 TurboCTC NC is supported natively in transformers. Until the next release, install it from source:

pip install git+https://github.com/huggingface/transformers.git datasets

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

model_id = "ibm-granite/granite-speech-5.0-470m-turboctc-nc"
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))

Model Architecture:

The architecture of granite-speech-5.0-470m-turboctc-nc consists of 16 conformer blocks trained with Connectionist Temporal Classification (CTC) with a 16,384 BPE classification head (see configuration below). We perform temporal subsampling by a factor of 8 to reduce frame rates from 100Hz to 12.5Hz: first by stacking and skipping consecutive logmel+delta frames (2x) followed by strided convolutions and pooled residuals in the first two conformer blocks (4x) as shown in the figure below. In addition, the encoder uses block-attention with blocks of 128 frames and self-conditioned CTC from the middle layer.

| Configuration parameter | Value | |-----------------|----------------------| | Input dimension | 320 (80 logmels + 80 deltas) x 2 | | Nb. of layers | 16 | | Hidden dimension | 1024 | | Nb. of attention heads | 8 | | Attention head size | 128 | | Attention block size | 128 | | Convolution kernel size | 7 | | Output dimension (BPE) | 16384 |

!conformer_blocks

Training Data:

Our training data is entirely comprised of publicly available datasets or of synthetic data generated from public corpora specifically targeting English ASR. A detailed description of the training datasets can be found in the table below:

| Name | Nb. hours | Source | |-----------|----------------|--------------| | CommonVoice-17 | 2500 | https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0 | | MLS | 44600 | https://huggingface.co/datasets/facebook/multilingual_librispeech | | Librispeech | 960 | https://huggingface.co/datasets/openslr/librispeech_asr | | VoxPopuli | 500 | https://huggingface.co/datasets/facebook/voxpopuli | | YODAS | 8900 | https://huggingface.co/datasets/espnet/yodas | | AMI | 150 | https://huggingface.co/datasets/edinburghcstr/ami | | Earnings-22 | 100 | https://huggingface.co/datasets/esb/datasets | | GigaSpeech | 10000 | https://huggingface.co/datasets/speechcolab/gigaspeech | | SPGI Speech | 4900 | https://huggingface.co/datasets/kensho/spgispeech |

In addition, the model was trained on three synthetic datasets:

1. 2000 hours of multi-speaker data generated by concatenating single-speaker segments from MLS, YODAS, CommonVoice-17, VoxPopuli, and AMI; 2. 500 hours of multi-speaker data generated by concatenating single-speaker segments from Earnings-22; and 3. 240 hours of utterances containing numbers, currencies, website names, phone numbers, addresses, and items containing decimal points or dots which were generated using either gpt-oss-120b or gpt-oss-20b and synthesized using StyleTTS2.

Infrastructure: We train Granite Speech TurboCTC NC using IBM's super computing cluster, Blue Vela, which is outfitted with NVIDIA H100 GPUs. This cluster provides a scalable and efficient...