OWSM v4

OWSM v4

Transcribes speech, translates it between languages, identifies the spoken language and decodes long recordings with the OWSM v4 model, with an optional text prompt. It suits speech researchers and anyone testing multilingual audio.

Made by espnet on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: OWSM v4 is a speech model demo from the ESPnet project. It takes 151 languages in and gives 25 translation targets out, and you can steer it with a text prompt. It can transcribe speech, translate it from any language to any other, name the spoken language and decode long recordings. The Space runs the owsm_v4_medium_1B model on ZeroGPU hardware.

Who it suits: It suits speech researchers, developers and anyone with multilingual recordings who wants to test an open speech model.

Industry | Who uses it | How they use it | Example
Speech research | Researchers | Compare an encoder-decoder model with the CTC demo on the same audio | A researcher runs one clip through both demos.
Media and archives | Archivists | Transcribe recordings in many languages | An archivist uploads an interview and reads the transcript.
Localization | Translators | Get a first draft translation of spoken audio | A translator picks a target language for a short recording.
Software | Developers | Read how a demo loads a model and asks for GPU time | A developer opens app.py to see the device choice.

If your recordings run longer than two minutes, the hosted demo will not take them, but a local copy can.

Where to use it:

• A web browser, through the Space's page on Hugging Face
• A local copy on Windows, macOS or Linux, using Python 3.12

Use case | Who | What they get back
Speech recognition | Anyone | A transcript of the audio
Speech translation | Translators | The speech rendered in one of the 25 target languages
Language identification | Researchers | The language the speaker used
Long recordings | Developers | Decoding of audio longer than the hosted limit, run locally

How to set it up:

• Python 3.12, since the project needs 3.12 or 3.13
• Python packages from requirements.txt
• A GPU for speed, though the DEVICE setting can force CPU

1. Install the huggingface_hub package.
2. Download the Space into a local folder named owsm-v4.
3. Move into that folder.
4. Install gradio at version 6.27.0.
5. Install the packages listed in requirements.txt.
6. Start the app.

pip install -U huggingface_hub
hf download espnet/owsm-v4 --repo-type space --local-dir owsm-v4
cd owsm-v4
pip install gradio==6.27.0
pip install -r requirements.txt
python app.py

How to use it: Open the Space's page, or start the app locally with the command below. Upload or record audio, pick a task and, if you like, add a text prompt. The hosted demo takes up to two minutes of audio.

python app.py

Tool | What it does
Speech recognition | Turns speech into text
Any-to-any translation | Renders speech in another language
Language identification | Detects which language is spoken
Long-form decoding | Handles recordings longer than one window
Text prompt | Primes the decoder, which this model supports

The repository also holds a second demo, OWSM-CTC v4, so you can compare the two on the same audio.

Aspect | This demo | The CTC demo
Model | owsm_v4_medium_1B, encoder-decoder | owsm_ctc_v4_1B, encoder-only
Decoding | Beam search | One encoder pass per 30 second window
Text prompt | Supported | Not supported

Example: You upload a recording of about a minute, choose speech recognition and leave the prompt empty. The demo returns the transcript. You then switch the task to translation and pick a target language.

For newcomers and experts: If you are new, upload a short clip and run speech recognition first. That shows what the model hears before you try translation or prompts.

If you know the field, the demo page comes from the ESPnet demo module, which reads the language menu, the translation targets and the prompt option off the checkpoint. This model has a decoder that a prompt can steer, which is why the option shows up here. The method is described in an Interspeech 2025 paper.

Upload or record audio
Pick a task and add a prompt
The model runs on a ZeroGPU slice
Text or translation comes back

Title: Language coverage
Languages in: 151 languages
Translation targets out: 25 languages

Good to know:

• The hosted demo accepts up to two minutes of audio, though the model has no such limit
• Setting DEVICE=cpu forces a CPU run, even on a machine with a GPU
• License: apache-2
• Cost to run: ZeroGPU hardware on Hugging Face for the hosted demo, and none named for a local run
• Data it reaches: the audio you give it
• Not for: recordings over two minutes on the hosted demo

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/espnet/owsm-v4
• Files: https://huggingface.co/spaces/espnet/owsm-v4/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.