Bayanat Labs Arabic ASR Channel Pairs
How much does a telephone line, a codec or an echoey room hurt Arabic speech recognition? This test set answers it with paired audio: read-speech recordings from FLEURS Egyptian Arabic, each released clean and rendered under fixed, documented channel conditions: a telephone band-limit, down-sampling, a mu-law telephone codec, low-bitrate Opus and synthetic room reverberation. Every variant keeps its parent's reference transcript and is aligned and level-matched to it, so any change in word error rate comes from the channel. The release includes the renderer, the conditions file, synthetic room impulse responses, and word error rates for two open Whisper models scored with ArabicWER. The conditions are simulated on read speech; they are not real telephone or call-centre recordings.
| Licence | CC BY 4.0 (audio, metadata and measurements); CC BY-SA 4.0 (transcript text); CC0 1.0 (conditions file and impulse responses); Apache-2.0 (renderer code) |
|---|---|
| Version | v0.1.0 · Dataset and renderer v0.1.0; the renderer is tagged on GitHub and not on PyPI yet. |
| Published | |
| Updated | |
| Who it is for | Speech-recognition teams that need to measure how much telephone, codec and room conditions raise Arabic word error rate, with clean and degraded audio of the same utterances. |
| Level | Intermediate |
Key numbers
Counted by script from the released files. Each key number is read from MANIFEST.json in the Hugging Face dataset and pinned to its SHA-256; the checksum file in the dataset lists the SHA-256 of every file.
Downloads and identifiers
| Where | What |
|---|---|
| Hugging Face | Dataset on Hugging Face |
| GitHub | Renderer on GitHub (arabic-channel-sim) |
| GitHub | Conditions file |
| bayanatlabs.com | ArabicWER, the scorer used for the tables |
Get started
Load a clean config and a channel config with the datasets library; their rows line up by parent_id. To render your own audio under the same conditions, install the renderer from GitHub at the v0.1.0 tag (pip install arabic-channel-sim will work once it is published on PyPI); the codec conditions need ffmpeg with libopus.
from datasets import load_dataset
clean = load_dataset("Bayanat-Labs/arabic-asr-channel-pairs", "CLEAN", split="test")
phone = load_dataset("Bayanat-Labs/arabic-asr-channel-pairs", "CD1", split="test")
# rows line up by parent_id
pip install "git+https://github.com/Bayanat-Labs/arabic-channel-sim@v0.1.0"
arabic-channel-sim render --condition CD1 --input clean.wav --output phone.flacWhat is released
| Part | Licence | In this release |
|---|---|---|
| Audio, clean and rendered (derived from FLEURS) | CC BY 4.0 | Yes |
| Metadata, measurements and word error rate tables | CC BY 4.0 | Yes |
| Transcript text: FLoRes-101 reference sentences and the model transcripts | CC BY-SA 4.0 | Yes |
| Conditions file and synthetic room impulse responses | CC0 1.0 | Yes |
| Renderer code (arabic-channel-sim) | Apache-2.0 | Yes |
Attribution: Arabic ASR Channel Pairs by Bayanat Labs, CC BY 4.0. Audio derived from Google FLEURS (CC BY 4.0); sentences from FLoRes-101 (CC BY-SA 4.0). https://huggingface.co/datasets/Bayanat-Labs/arabic-asr-channel-pairs
Intended uses
- Measure how much each channel condition raises a speech recogniser's word error rate on the same Arabic utterances, and report the change against clean audio.
- Compare speech recognisers on robustness to telephone band-limiting, codecs and reverberation.
- Render your own clean audio under the same documented conditions with the renderer.
Out of scope
- Claims about real calls: the conditions are simulations on read speech, without noise, packet loss, echo cancellation or handset acoustics.
- Training on the test split: it is meant for evaluation.
- Speaker or gender analysis: FLEURS publishes no speaker ids, and every selected recording is labelled female.
Rights at a glance
Training, display, redistribution and synthetic-voice use are separate permissions.
| Use | Status | Note |
|---|---|---|
| Training | Allowed | Audio, metadata and measurements under CC BY 4.0 and transcript text under CC BY-SA 4.0, with attribution. This is a test set: training on it invalidates comparisons made with it. |
| Display | Allowed | With attribution: audio derived from Google FLEURS (CC BY 4.0), sentences from FLoRes-101 (CC BY-SA 4.0), channel rendering by Bayanat Labs. |
| Redistribution | Allowed | Under each part's licence: modified transcript text stays under CC BY-SA 4.0; conditions file and impulse responses CC0 1.0; renderer code Apache-2.0. |
| Synthetic voice | Restricted | Not cleared for building voices that imitate the speakers: the FLEURS licence covers the recordings, not the speakers' consent to voice cloning. |
Limitations
- Read Modern Standard Arabic by Egyptian speakers; not conversational and not dialectal.
- The word error rate tables cover a subset of the test recordings and two open Whisper models; their intervals are wide, and differences of a point or two between conditions are within noise.
- The publisher of Whisper has not listed its training data, so read the change against clean audio rather than the absolute scores.
- Install the renderer from GitHub for now; it is not on PyPI yet.
Sources and licences
| Source | Licence | Redistributed here |
|---|---|---|
| FLEURS (google/fleurs), Egyptian Arabic (ar_eg) | CC BY 4.0 (Google) | Yes |
| FLoRes-101 sentences (read aloud in FLEURS) | CC BY-SA 4.0 | Yes |
| Whisper models (openai/whisper-large-v3-turbo and openai/whisper-small), run to produce the transcripts | MIT | No |
Channel rendering, alignment, measurements and word error rate tables are by Bayanat Labs. The conditions file and the impulse responses, generated with pyroomacoustics, are dedicated to the public domain under CC0 1.0. The underlying recordings are not owned by Bayanat Labs. The transcripts were produced by Bayanat Labs with two open Whisper models and scored with ArabicWER. Source details
How to cite
Cite as:
Bayanat Labs (2026). Arabic ASR Channel Pairs, version 0.1.0. Bayanat Labs. https://huggingface.co/datasets/Bayanat-Labs/arabic-asr-channel-pairs
No DOI. Cite the URL in the citation above and the date you read it.
@misc{bayanatlabs2026arabicasrchannelpairs,
author = {{Bayanat Labs}},
title = {Arabic ASR Channel Pairs, version 0.1.0},
year = {2026},
month = oct,
howpublished = {\url{https://huggingface.co/datasets/Bayanat-Labs/arabic-asr-channel-pairs}}
}
Changelog
| Date | Version | Change |
|---|---|---|
| v0.1.0 | Clean and rendered audio, metadata, measurements and Whisper word error rate tables on Hugging Face; the renderer, conditions file and impulse responses on GitHub. |
Corrections and removal requests
If something in this release should not be there, or is wrong, email hello@bayanatlabs.com. We handle removal requests within 30 days, and platform requests within 24 hours. How removal requests work
Related product: Bayanat Voice