Granite-Speech-5.0-470M-TurboCTC

πŸŽ‰ NEW (October 21, 2026): Granite Speech 5.0 TurboCTC is now multilingual!
🌍 7 Languages (En, Fr, De, Es, Pt, It, Ja)
❓ Native punctuation & capitalization
πŸš€ Better English WER (5.51% vs. 5.78%)
βͺ Previous English-only release: revision 947f59a

Model Summary: Granite Speech 5.0 TurboCTC is a compact 470 million parameter multilingual ASR model for English, French, German, Spanish, Portuguese, Italian and Japanese with very high inference speed that is well suited for deployment on laptops, smartphones and other edge devices. 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 multilingual BPE units. It was trained on approximately 153,000 hours of multilingual audio from public corpora using Connectionist Temporal Classification (CTC) and inference is done non-autoregressively with greedy decoding. The model outputs punctuated and capitalized transcripts.

Evaluations:

We evaluated granite-speech-5.0-470m-turboctc on standard short-form English ASR benchmarks from the Open ASR leaderboard. Despite adding six languages and native punctuation, the multilingual update improves the average English WER over the original English-only release (5.51% vs. 5.78%):
granite-speech-5.0-470m-turboctc-wer-en

We also evaluated multilingual ASR on the Open ASR leaderboard multilingual benchmarks (FLEURS, CommonVoice, MLS) for German, Spanish, French, Italian and Portuguese, against the best open-weights model of each organization:
granite-speech-5.0-470m-turboctc-wer-ml

For Japanese, which is not part of the Open ASR leaderboard multilingual benchmarks, we report the character error rate (CER) on CSJ eval1, CommonVoice-17 and FLEURS. The 470M-parameter granite-speech-5.0-470m-turboctc outperforms the 2B-parameter granite-speech-5.0-2b on all three test sets:
granite-speech-5.0-470m-turboctc-cer-ja

Accuracy vs. speed on the Open ASR leaderboard (English short-form, average over 9 public test sets, official results as of October 21, 2026; the orange marker is this model's unofficial result, the leaderboard entry is still the English-only release):
granite-speech-5.0-470m-turboctc-pareto-rtfx

Release Date: October 21, 2026 (original English-only release: August 25, 2026)

License: Apache 2.0

Supported Languages: English, French, German, Spanish, Portuguese, Italian, Japanese

Intended Use: The model is intended to be used in enterprise applications that involve accurate low-latency/high-throughput multilingual speech-to-text transcription.

Usage:

Usage with transformers

Granite Speech 5.0 TurboCTC is supported natively in transformers>=5.16.0:

pip install transformers>=5.16.0 datasets
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))
Usage with mlx-audio for Apple Silicon M series chips

Install a recent version of mlx-audio (0.5.1 or later):

pip install -U mlx-audio

Sample use:

python -m mlx_audio.stt.generate --model ibm-granite/granite-speech-5.0-470m-turboctc --verbose --audio "audio.wav" --output-path "transcript"
Usage with transcribe.cpp for local GGUF inference

transcribe.cpp runs the model from a quantized GGUF on Metal, Vulkan, CUDA, ROCm, or CPU, with no Python at runtime. Build it (tested with release 0.4.1), then fetch a pre-quantized GGUF of this multilingual release from vladers/granite-speech-5.0-470m-turboctc-GGUF (community conversion; Q8_0 recommended, Q6_K to Q4_K_M also available):

git clone --branch v0.4.1 https://github.com/handy-computer/transcribe.cpp
cd transcribe.cpp && cmake -B build && cmake --build build --target transcribe-cli
hf download vladers/granite-speech-5.0-470m-turboctc-GGUF \
  granite-speech-5.0-470m-turboctc-Q8_0.gguf --local-dir models

Sample use (16 kHz mono WAV; -o writes only the transcript to a file):

build/bin/transcribe-cli -q -m models/granite-speech-5.0-470m-turboctc-Q8_0.gguf -o transcript.txt audio.wav

On silence or non-speech audio, transcribe.cpp returns the model's <|nospeech|> control token as the transcript; treat that output as no speech.

Model Architecture:

The architecture of granite-speech-5.0-470m-turboctc is unchanged from the original English-only release, apart from the multilingual output vocabulary, and consists of 16 conformer blocks trained with Connectionist Temporal Classification (CTC) with a 16,384 multilingual 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 (multilingual BPE) 16384

conformer_blocks

Training Data:

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

Name Nb. hours Source
CommonVoice-17 En,De,Es,Fr,It,Pt,Ja 5,660 https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0
MLS English (subset) 31,640 https://huggingface.co/datasets/facebook/multilingual_librispeech
MLS De,Es,Fr,It,Pt 3,890 https://huggingface.co/datasets/facebook/multilingual_librispeech
Librispeech-PC English 950 https://huggingface.co/datasets/yoom618/librispeech_pc
LibriHeavy Large English 45,750 https://huggingface.co/datasets/anyspeech/libri-heavy
VoxPopuli En,De,Es,Fr,It 1,170 https://huggingface.co/datasets/facebook/voxpopuli
VoxPopuli Granary English 23,730 https://huggingface.co/datasets/nvidia/Granary
YODAS English 1,230 https://huggingface.co/datasets/espnet/yodas
YODAS Granary English (filler-only segments removed) 16,220 https://huggingface.co/datasets/nvidia/Granary
YODAS Granary Spanish 2,290 https://huggingface.co/datasets/nvidia/Granary
YODAS Granary French 1,650 https://huggingface.co/datasets/nvidia/Granary
YODAS Granary German 2,130 https://huggingface.co/datasets/nvidia/Granary
YODAS Granary Portuguese 1,170 https://huggingface.co/datasets/nvidia/Granary
YODAS Granary Italian 770 https://huggingface.co/datasets/nvidia/Granary
AMI English 150 https://huggingface.co/datasets/edinburghcstr/ami
Earnings-22 English 105 https://huggingface.co/datasets/esb/datasets
YODAS Japanese 1,420 https://huggingface.co/datasets/espnet/yodas
ReazonSpeech Japanese (filtered) 8,130 https://huggingface.co/datasets/reazon-research/reazonspeech
Multi-speaker mixtures English (synthetic) 2,500 Internal: overlapped-speech mixtures built from AMI, CommonVoice and VoxPopuli
Written-form TTS English (numbers, codes, URLs; 16 kHz + 8 kHz) 240 Internal TTS
CommonVoice-17 Japanese paraphrase TTS 130 Internal TTS of paraphrased CommonVoice-17 Japanese text
Granite-text TTS Japanese 1,860 Internal TTS of Granite-generated Japanese text
MUSAN noise + music (non-speech, empty target) 43 https://www.openslr.org/17/

Infrastructure: We train Granite Speech TurboCTC using IBM's super computing cluster, Blue Vela, which is outfitted with NVIDIA H100 GPUs. This cluster provides a scalable and efficient infrastructure for training our models over thousands of GPUs. The training of this particular model was completed in 24 days on 8 H100 GPUs.

Resources

Citation

@misc{granite-speech-5.0-470m-turboctc,
  title={Granite Speech 5.0 TurboCTC},
  author={IBM Granite Speech Team},
  year={2026},
  url={https://huggingface.co/ibm-granite/granite-speech-5.0-470m-turboctc}
}
Downloads last month
51,699
Safetensors
Model size
0.5B params
Tensor type
F32
Β·
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ 1 Ask for provider support

Model tree for ibm-granite/granite-speech-5.0-470m-turboctc

Finetunes
7 models
Quantizations
25 models

Spaces using ibm-granite/granite-speech-5.0-470m-turboctc 6

Collection including ibm-granite/granite-speech-5.0-470m-turboctc

Paper for ibm-granite/granite-speech-5.0-470m-turboctc

Articles mentioning ibm-granite/granite-speech-5.0-470m-turboctc