IPA Transcription Leaderboard

IPA Transcription Leaderboard

Ranks models that turn English speech into IPA phonetic symbols, scored on native, non-native and impaired speech. Lets model authors submit their own model and follow its evaluation.

Made by KoelLabs 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: IPA Transcription Leaderboard compares models that turn English speech into IPA, the phonetic alphabet. Koel Labs scores the models on three kinds of English speech, from native speakers, from non-native speakers and from people whose speech is impaired. The ranking uses two scores, one for how often the phonemes are wrong and one for how often the phonetic features are wrong. You can also submit your own model to be scored.

Who it suits: Speech researchers, model authors, and developers who build pronunciation feedback or speech tools that must cope with different accents and speech conditions.

Industry | Who uses it | How they use it | Example
Speech technology | Speech researchers | Compare phonetic transcription models on one scoreboard | Sort the table by PER or FER
Language learning | App developers | Pick a model for pronunciation feedback | Check scores on non-native speech
Accessibility | Speech tool developers | See how models handle impaired speech | Look at the dataset-level FER comparisons
Machine learning | Model authors | Submit their own model to be scored | Enter a model ID in the Submit Model tab

Anyone who does not work with speech models can skip it.

Where to use it: It runs in these places.

• A web browser, on the Space's page at Hugging Face
• Windows, with Python 3.11
• macOS, with Python 3.11
• Linux, with Python 3.11

Use case | Who | What they get back
Compare models | Researchers | A ranked table of PER and FER scores, where lower is better
Submit a model | Model authors | An evaluation run for the model
Check progress | Model authors | Updates on the evaluation in the status tab

How to set it up: Using the Space online needs only a browser. A copy on your own computer needs the following.

• Python 3.11 and pip
• The huggingface_hub package, which gives you the hf command
• Disk space for the repository, which holds 239 files

1. Install the huggingface_hub package.
2. Download the Space's files into a local folder.
3. Move into that folder.
4. Install Gradio at the version the Space uses, then install the packages in requirements.txt.
5. Start the app from the app folder.

pip install -U huggingface_hub
hf download KoelLabs/IPA-Transcription-EN --repo-type space --local-dir IPA-Transcription-EN
cd IPA-Transcription-EN
pip install gradio==5.12.0
pip install -r requirements.txt
python app/app.py

How to use it: Open the leaderboard to read and sort the scores. To submit a model, open the Submit Model tab and fill in the details. Then follow the run in the Submission Status tab.

Tool | What it does
Leaderboard | Shows PER and FER scores, sortable by FER, PER or model size
Submit Model tab | Takes a model ID, a submission name and an optional link
Submission Status tab | Shows live progress when you enter a model ID or task ID
DEVELOPMENT.md | The repository's notes for running the leaderboard locally

Example: You open the Submit Model tab, enter your phoneme model's ID and a submission name, and click Submit. Then you enter the model ID in the Submission Status tab and watch the evaluation. It can take several hours.

For newcomers and experts: IPA is a set of symbols for speech sounds, and a phoneme is one sound unit. PER counts how often a model gets those sounds wrong, and lower is better. The first small thing to try is opening the leaderboard and sorting by PER.

FER is the finer measure, because it also looks at the phonemic features of each sound. Scoring uses shared IPA spelling normalization and full feature-parser coverage. The Space names six transcription models and two datasets, one of scores and one of test data. Sorting by size lists the smallest models first and puts unknown sizes last.

You submit a model ID
The model is queued for evaluation
It is scored on English speech sets
PER and FER appear on the leaderboard

Good to know: Evaluation is slow, and the Space runs on a basic compute plan with 16 GB of RAM and 2 vCPUs.

• License: agpl-3, the GNU Affero General Public License
• Cost to run: none named
• Data it reaches: the model details you submit, plus the score and test datasets the Space names
• Not for: reusing the Koel Labs brand, logos, blog posts or site content, because the maker keeps those rights
• Hardware: the Space is set to cpu-basic at Hugging Face

On Hugging Face, 13 November 2026

• Likes: 10
• Built with: gradio
• License: agpl-3.0

Links

• Space: https://huggingface.co/spaces/KoelLabs/IPA-Transcription-EN
• Files: https://huggingface.co/spaces/KoelLabs/IPA-Transcription-EN/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.