README.md

metadata

title: OWSM-CTC v4
emoji: 🎙️
colorFrom: blue
colorTo: green
sdk: gradio
python_version: '3.12'
sdk_version: 6.27.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: 151 languages in, 25 translation targets out, fast
tags:
  - espnet
  - owsm
  - automatic-speech-recognition
  - speech-translation
  - language-identification
models:
  - espnet/owsm_ctc_v4_1B

OWSM-CTC v4

The source of espnet/owsm-ctc-v4, one of two OWSM v4 demos kept in this repository. Both offer the same tasks - speech recognition, any-to-any speech translation, language identification and long-form decoding - so the two models can be compared on the same audio:

	this demo	espnet/owsm-v4
model	owsm_ctc_v4_1B, encoder-only	owsm_v4_medium_1B, encoder-decoder
decoding	one encoder pass per 30 s window, no beam search	beam search
text prompt	not supported by OWSM-CTC	supported
source	this directory	egs2/owsm_v4/s2t1/demo

The language menu and the translation targets are read from the checkpoint, so a model covering more languages needs no edit here.

The hosted demo takes audio of up to two minutes and asks ZeroGPU for a matching slice of GPU time; the model itself has no such limit, so run the app yourself for longer recordings.

The card pins python_version: "3.12". espnet requires 3.12 or 3.13, and a Space image with an older interpreter installs no espnet at all - pip reports "No matching distribution found" and the build fails.

Running and publishing

pip install -r requirements.txt gradio
python app.py                     # http://127.0.0.1:7860

DEVICE decides where the models run and takes precedence over everything else, so DEVICE=cpu is a local CPU test even on a machine with a GPU. With DEVICE unset, the choice is CUDA when the runtime is ZeroGPU (SPACES_ZERO_GPU, the marker Hugging Face sets — asking torch.cuda.is_available() there answers False, because the GPU is attached only while a @spaces.GPU function runs) or when torch reports a GPU, and CPU otherwise.

To publish, from the ESPnet checkout:

hf auth login
hf upload espnet/owsm-ctc-v4 egs2/owsm_ctc_v4/s2t1/demo . --repo-type space

Not before the release named in requirements.txt. A Space installs espnet from PyPI, so an app that calls something only master has is a Space that builds and then fails to start. This app decodes with best_path() and the decode_long() that takes a path and a batch size, which arrived after 202610 - hence espnet>=202610.post1, which pip will refuse until that release exists. Uploading this directory before then replaces a working Space with a broken one, which is exactly what happened on 2026-09-20.

Citation

@inproceedings{owsm-ctc,
  title={{OWSM-CTC}: An Open Encoder-Only Speech Foundation Model for Speech
         Recognition, Translation, and Language Identification},
  author={Yifan Peng and Yui Sudo and Muhammad Shakeel and Shinji Watanabe},
  booktitle={Proc. ACL},
  year={2024}
}