ESPNet2 ASR Librispeech word vs bpe tokens

ESPNet2 ASR Librispeech word vs bpe tokens

Compares three English speech recognition models from one ESPnet2 Conformer setup, using bpe, character and word tokens, in a Gradio page with example audio. For speech researchers and students.

Made by jkang 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: This Gradio app puts three English speech recognition models side by side. They are ESPnet2 Conformer models trained on Librispeech, and the author's notes say they differ in how text is split: bpe, character or word units. The repository holds example audio in two folders.

Who it suits: It suits speech researchers and students who want to hear how the choice of token unit changes a recognizer's output. Developers can read the main script as a small multi-model demo. The record gives no error rates.

Industry | Who uses it | How they use it | Example
Research | Speech researchers | Compare token units for the same kind of model | Run one clip through all three models
Education | Students of speech technology | See how bpe, character and word outputs differ | Compare the three transcripts of one clip
Software | Python developers | Read a Gradio script that serves three models | Open the main file and packages.txt

You can skip it if your audio is not in English.

Where to use it:

• A web browser, on the Space's page on Hugging Face
• Windows, macOS or Linux with Python, from a downloaded copy

Use case | Who | What they get back
Compare token units | Researchers | Transcripts from the bpe, character and word models
Try the bundled examples | Students | A quick sense of how each model hears
Study the script | Developers | A Gradio app that serves several models

How to set it up:

• Nothing at all if you use the online Space
• Python on your computer, only if you run a copy
• The packages listed in requirements.txt
• System packages from packages.txt, which Hugging Face installs with apt-get

1. Open the Space's page in a browser to use it with nothing to install.
2. To run a copy, install the Hugging Face Hub package and download the Space.
3. Move into the new folder, then install Gradio and the listed requirements.
4. Start the app with Python, using the main file named in the command below.

pip install -U huggingface_hub
hf download jkang/espnet2_librispeech_100h_word_vs_bpe_vs_char --repo-type space --local-dir espnet2_librispeech_100h_word_vs_bpe_vs_char
cd espnet2_librispeech_100h_word_vs_bpe_vs_char
pip install gradio
pip install -r requirements.txt
python gradio_asr_en_libri100_word_vs_bpe.py

How to use it: Online, open the Space and use the controls on its page. On your own machine, run the main file from the Space folder, then use the page it serves. The example audio files are the quickest way to try it.

python gradio_asr_en_libri100_word_vs_bpe.py

Example: You open the Space and run one of the example clips. The app returns the English text from each of the three models, so you can compare them line by line. The record does not describe the controls, so check the page.

For newcomers and experts: A speech recognizer writes text in small pieces. Those pieces can be whole words, single characters or word parts, called bpe. Start by opening the Space and running one example to see how the three outputs differ.

If you know the field, the three models share one Conformer setup and differ in their token unit, which makes this a handy side-by-side view of that choice. The author's notes mention audio reading problems with live recording, so test with the examples first.

You choose an example or give audio
The Gradio page sends it to the script
Three ESPnet2 models decode it
The page shows each transcript

Good to know:

• License: MIT, as listed on the Space
• Cost to run: none named
• Data it reaches: the audio you give it, the three models it names and the example clips
• Not for: non-English speech, and live recording may not work, as the author's notes say

On Hugging Face, 13 November 2026

• Likes: 3
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/jkang/espnet2_librispeech_100h_word_vs_bpe_vs_char
• Files: https://huggingface.co/spaces/jkang/espnet2_librispeech_100h_word_vs_bpe_vs_char/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.