Turkish ASR Leaderboard

Turkish ASR Leaderboard

Compares Turkish speech-to-text models by their benchmark and timing results on a Google test dataset. It helps you choose a speech recognition model for Turkish audio.

Made by EmreAkgul 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 Space compares Turkish speech-to-text models on the Turkish test portion of Google's FLEURS dataset. It lays out benchmark results and timing results, so you can weigh the models against each other before choosing one. Emre Akgul maintains it, and it draws on his separate evaluation project.

Who it suits: It suits anyone choosing a speech recognition model for Turkish audio, and researchers who track how such models compare. Developers adding voice features for Turkish speakers can read the timing results beside the benchmark results.

Industry | Who uses it | How they use it | Example
Research | ASR researchers | Compare benchmark results across Turkish models | See where a newer model sits against the ones already listed
Software | Developers adding voice input | Read timing results beside benchmark results | Pick a model that fits an app's speed needs
Media | Teams that transcribe Turkish audio | Shortlist models to try on their own recordings | Narrow a long list before running a trial
Education | Students of speech technology | Study how Turkish models differ | Use a public comparison as a worked example

If you work with languages other than Turkish, the results will not help you.

Where to use it: The leaderboard runs in these places.

• A web browser: open the Space page on Hugging Face, with nothing to install.
• Windows, macOS or Linux with Python: run a copy of the app on your own computer.
• A local browser tab: the app prints a local address when it starts, and you open that address.

Use case | Who | What they get back
Compare models | Researchers | Benchmark results for Turkish speech-to-text models on one shared test set
Check speed | Developers | Timing results for the same models
Study or adapt the code | People who build leaderboards for other languages | A Gradio app and CSV files they can read and change under the Apache license

How to set it up: You need a few things first.

• Python, with version 3.10 used in the project's own notes for local runs
• pip, to install packages
• A terminal on Windows, macOS or Linux

These are Hugging Face's steps for a local copy.

1. Install the Hugging Face hub package with pip.
2. Download the Space's files into a folder named Turkish-ASR-Leaderboard.
3. Move into that folder.
4. Install Gradio 6.28.0, then the packages listed in requirements.txt.
5. Start the app.

pip install -U huggingface_hub
hf download EmreAkgul/Turkish-ASR-Leaderboard --repo-type space --local-dir Turkish-ASR-Leaderboard
cd Turkish-ASR-Leaderboard
pip install gradio==6.28.0
pip install -r requirements.txt
python app.py

The project's own notes describe another route, with a virtual environment and make.

python3.10 -m venv .venv
source .venv/bin/activate
make install
make run

How to use it: Open the Space page, or start your local copy with the command below and open the local address Gradio prints in the terminal. The page then shows benchmark and timing results for the Turkish models it covers.

python app.py

Tool | What it does
python app.py | Starts the Gradio app from your downloaded copy
make run | Starts the app after make install, in the project's own route

Example: You are picking a model to caption Turkish videos. You open the leaderboard, read the benchmark results for the models listed there, then check their timing results. You leave with a short list of models worth testing on your own recordings.

For newcomers and experts: A speech-to-text model turns spoken words into written text, and a leaderboard lets you compare several of them on the same recordings. Start by opening the Space page in your browser and reading the benchmark results. Then look at the timing results to see how the models differ.

The repository keeps its data in benchmark.csv, models.csv, samples.csv and timing.csv beside app.py, with more code under src/. A requirements.txt file keeps the Gradio version in step with the runtime that Hugging Face provides. All scores come from one test set, so they may not carry over to your own audio. The code is adapted from a Hebrew transcription leaderboard by ivrit-ai.

Models are evaluated on the FLEURS Turkish test data
Benchmark and timing results are kept in CSV files
app.py builds the Gradio page
You read the leaderboard in a browser

Good to know: A few facts are worth checking first.

• Python: the project's notes use Python 3.10 for local runs.
• License: apache-2, as the Space's listing shows. The notes say the code was adapted from an Apache 2.0 project.
• Cost to run: none named. The source lists no API key, account or paid service.
• Data it reaches: downloading a copy pulls the Space's files from Hugging Face into a folder you choose. The source names no other account or service.
• Not for: other languages, since the results cover Turkish models only.

On Hugging Face, 13 November 2026

• Likes: 10
• Built with: gradio
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/EmreAkgul/Turkish-ASR-Leaderboard
• Files: https://huggingface.co/spaces/EmreAkgul/Turkish-ASR-Leaderboard/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.