futo-org/gravel-ctc-440m

gravel-ctc-440m

A 440M CTC English ASR model: the conformer speech encoder of `ibm-granite/granite-speech-4.1-2b-nar`, extracted, LoRA-finetuned with a BPE-256 CTC head, and merged into a standalone greedy-CTC recognizer.

[!NOTE] Gravel is granite broken into smaller pieces. This model is the 440M encoder chipped out of IBM's 2B+ `ibm-granite/granite-speech-4.1-2b-nar` plus a small CTC head. It is not an IBM release, it is a derivative of the granite model.

What it is

• —Encoder: the granite_speech conformer encoder from ibm-granite/granite-speech-4.1-2b-nar instantiated from transformers.models.granite_speech classes, weights extracted, then LoRA-adapted (r=16 on attention + macaron-FFN linears) and merged.
• —Head: a new BPE-256 head (257 with blank) behind a 50→25 Hz average-pool-initialized depthwise downsample.
• —Frontend: log-mel (n_fft 512, win 400, hop 160, 80 mels) → Whisper-style per-utterance norm → 2-frame stack (160-dim @ 50 Hz), computed inside the model — feed raw 16 kHz PCM.

How it was made

1. 1.Extract the encoder weights + config from ibm-granite/granite-speech-4.1-2b-nar.
2. 2.Attach the 50→25 Hz downsample + new BPE-256 CTC head (the vocab is a 256-piece BPE)
3. 3.Train with CTC + an emission penalty (a from-scratch BPE-256 head blank-collapses without it) on a large English mix, backbone frozen, {LoRA, downsample, head} trainable.
4. 4.Merge LoRA → this standalone checkpoint.

It is used as the knowledge-distillation teacher for FUTO's streaming ASR models.

Usage

import torch, torchaudio
from transformers import AutoModel

model = AutoModel.from_pretrained("futo-org/gravel-ctc-440m", trust_remote_code=True).eval()

wav, sr = torchaudio.load("speech.wav")           # any mono/stereo file
wav = torchaudio.functional.resample(wav.mean(0), sr, 16000)
print(model.transcribe(wav))                       # greedy CTC text

Or via pipeline:

from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="futo-org/gravel-ctc-440m",
               trust_remote_code=True)
print(asr("speech.wav")["text"])

Lineage & license

Derived from `ibm-granite/granite-speech-4.1-2b-nar` (Apache-2.0, IBM).

Modifications: encoder extraction, 50→25 Hz downsample, BPE-256 CTC head, LoRA finetune (merged).