Sygma ASR Arabic

Sygma ASR Arabic

A test bench that records or takes Arabic speech and compares Sygma's decoding with the base model's greedy output, with word error rates. For speech researchers and developers.

Made by abedbanna 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: Sygma ASR Arabic test bench lets you record or upload Arabic speech and compare two ways of turning it into text. One is Sygma's own decoding, which uses an 8-best beam search with a word-count bonus and a breakdown filter. The other is plain greedy decoding from the base model, run on that same recording.

Who it suits: Speech researchers and developers who work on Arabic speech recognition, and anyone who wants to hear how the two decoders differ on their own clips.

Industry | Who uses it | How they use it | Example
Speech technology | Researchers | Compare decoding methods on the same audio | Running one clip and reading both transcripts
Software | Developers | Test Arabic transcription before building on it | Recording a short sentence and checking the text
Language work | Linguists | Look at how dialectal speech is transcribed | Uploading a clip of spoken Arabic

You can skip it if you need clips longer than 120 seconds or speech in other languages, since the demo handles Arabic clips up to that length.

Where to use it:

• A web browser, on the Space's own page at Hugging Face
• A copy on your own computer, using Python 3.10 on Windows, macOS or Linux

Use case | Who | What they get back
Compare decoders | Researchers | Two transcripts of one clip, one from each method
Score accuracy | Developers | A word error rate for each, when you paste a reference transcript

How to set it up:

• A browser and a microphone or an audio file, if you use the Space online
• Python 3.10 and pip, if you run a copy yourself
• Room for a model of about 8 GB, which downloads when the Space wakes up

1. Open the Space's page on Hugging Face to use it online. Nothing needs installing.
2. For a local copy, install the huggingface_hub package and download the Space's files.
3. Move into the new folder, then install Gradio and the packages in requirements.txt.
4. Start the app with python.

pip install -U huggingface_hub
hf download abedbanna/Sygma-ASR-Arabic-demo --repo-type space --local-dir Sygma-ASR-Arabic-demo
cd Sygma-ASR-Arabic-demo
pip install gradio==5.49.1
pip install -r requirements.txt

How to use it: Record Arabic audio or upload a clip of up to 120 seconds. Optionally paste a reference transcript, then run it to see both outputs. The first load takes a few minutes while the model downloads. On your own computer, start the app first and open the page it serves.

python app.py

Tool | What it does
Record or upload | Takes Arabic speech from a microphone or an audio file
Sygma decoding | Transcribes with the 8-best beam search, word-count bonus and breakdown filter
Greedy output | Transcribes the same clip with the base model's plain decoding
Reference transcript | Lets the page score each result with a word error rate

Example: You record a short Arabic sentence and paste what you said as the reference. The page shows the Sygma transcript and the greedy transcript, each with its word error rate.

For newcomers and experts: If you are new to this, speech recognition turns audio into text, and word error rate counts how many words came out wrong. Record one short sentence and paste in what you said to see both scores.

For experts, the page compares length-corrected n-best decoding against greedy decoding from the base model, CohereLabs/cohere-transcribe-arabic-07-2026. It runs on ZeroGPU, and the model comes down at wake-up, so a cold start is slow.

Record or upload Arabic audio
Both decoders run on the clip
Optional reference transcript is read
Word error rates are computed
Both transcripts show on the page

Good to know:

• Built with Gradio 5.49.1 and Python 3.10, and the main file is app.py
• License: apache-2
• Cost to run: none named, but the Space runs on ZeroGPU
• Data it reaches: the audio you record or upload, and the model files it downloads
• Not for: clips over 120 seconds

On Hugging Face, 13 November 2026

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

Links

• Space: https://huggingface.co/spaces/abedbanna/Sygma-ASR-Arabic-demo
• Files: https://huggingface.co/spaces/abedbanna/Sygma-ASR-Arabic-demo/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.