README.md

---
license: other
language:
- ar
task_categories:
- automatic-speech-recognition
pretty_name: Quranic ASR Provider Benchmark Results
size_categories:
- 1K<n<10K
tags:
- quran
- arabic
- speech-recognition
- benchmark
- asr-evaluation
- commercial-asr
- tajweed
configs:
- config_name: benchmark
  data_files:
  - split: test
    path: data/benchmark/test.jsonl
- config_name: hypotheses
  data_files:
  - split: test
    path: data/hypotheses/test.jsonl
- config_name: scores
  data_files:
  - split: test
    path: data/scores/test.jsonl
---

# Quranic ASR Provider Benchmark Results

Professional benchmark artifacts for comparing commercial and official ASR providers on the Quranic ASR benchmark hosted at `Quran-Lab/quranic-asr-benchmark`.

This repository contains metadata, normalized result tables, raw provider responses, unchanged run scripts, scoring outputs, Tarteel streaming probes, and reports. It does not duplicate the source audio.

## What Is Included

| Area | Path | Purpose |
| --- | --- | --- |
| Benchmark split | `data/benchmark/test.jsonl` | 600 clip IDs, references, source labels, and source-audio pointers |
| Hypotheses split | `data/hypotheses/test.jsonl` | Final transcription outputs for each provider/model/clip |
| Scores split | `data/scores/test.jsonl` | Official WER/CER metrics by provider and source |
| Raw outputs | `artifacts/outputs/raw/` | Raw API/websocket responses captured during runs |
| Provider scripts | `artifacts/scripts/` | Run harnesses copied byte-for-byte from the benchmark run |
| Official scorer | `artifacts/source_benchmark/score.py` | Source benchmark scorer used for all reported metrics |
| Reports | `artifacts/reports/` | Consolidated result notes and Tarteel behavior report |
| Checksums | `MANIFEST.sha256` | SHA-256 checksums for all uploaded files |

## What Is Not Included

Audio files are intentionally not included in this repository.

The `audio_path` column points to the corresponding files in `Quran-Lab/quranic-asr-benchmark`. Use that dataset for audio access, licensing, and upstream provenance.

No provider API keys, OAuth tokens, account passwords, local credential files, or auth tokens are included.

## Dataset Configs

This repository exposes three clean Hugging Face dataset configs.

| Config | Split | Rows | Description |
| --- | --- | ---: | --- |
| `benchmark` | `test` | 600 | Benchmark metadata and references |
| `hypotheses` | `test` | 3,600 | Six provider/model outputs for each clip |
| `scores` | `test` | 24 | Per-provider metrics for each source plus overall |

Example loading code:

```python
from datasets import load_dataset

benchmark = load_dataset("Quran-Lab/quranic-asr-cloud-rawdata", "benchmark")
hypotheses = load_dataset("Quran-Lab/quranic-asr-cloud-rawdata", "hypotheses")
scores = load_dataset("Quran-Lab/quranic-asr-cloud-rawdata", "scores")
```

## Schema

### `benchmark`

| Column | Type | Description |
| --- | --- | --- |
| `id` | string | Stable clip ID from the source benchmark |
| `source` | string | One of `everyayah_heldout`, `qul_alnufais`, `tlog_holdout` |
| `reference_text` | string | Quranic reference text used by the official scorer |
| `audio_dataset` | string | Source dataset containing the audio |
| `audio_path` | string | Relative audio path inside the source dataset |

### `hypotheses`

| Column | Type | Description |
| --- | --- | --- |
| `id` | string | Stable clip ID |
| `source` | string | Benchmark source subset |
| `provider` | string | ASR provider or official service |
| `model` | string | Provider model name |
| `run_type` | string | API mode used for the run |
| `hypothesis_text` | string | Final ASR transcript emitted by the provider |
| `has_hypothesis` | bool | Whether the provider returned non-empty text for the clip |
| `reference_text` | string | Reference text for convenience |
| `audio_dataset` | string | Source dataset containing the audio |
| `audio_path` | string | Relative audio path inside the source dataset |

### `scores`

| Column | Type | Description |
| --- | --- | --- |
| `provider` | string | ASR provider or official service |
| `model` | string | Provider model name |
| `run_type` | string | API mode used for the run |
| `source` | string | Source subset or `ALL` |
| `clips` | int | Number of clips scored |
| `wer` | float | Word error rate from the official scorer |
| `cer` | float | Character error rate from the official scorer |
| `wer_alef` | float | Alef-insensitive WER from the official scorer |

## Benchmark Setup

- Source benchmark: `Quran-Lab/quranic-asr-benchmark`
- Clips: 600 total, 200 per source
- Duration: 7,701.439 seconds, about 2.14 hours
- Sources: `everyayah_heldout`, `qul_alnufais`, `tlog_holdout`
- Audio format in source dataset: 16 kHz mono WAV
- Scorer: `artifacts/source_benchmark/score.py`
- Metrics: WER, CER, and alef-insensitive WER/CER

## Overall Results

| Rank | Provider / model | Run type | WER | CER | WER(alef) |
| ---: | --- | --- | ---: | ---: | ---: |
| 1 | Tarteel official | realtime websocket | 10.99 | 7.14 | 9.72 |
| 2 | Google Chirp 3 | sync recognize | 11.92 | 7.87 | 10.55 |
| 3 | Google Chirp 3 | realtime streaming recognize | 13.56 | 8.67 | 12.22 |
| 4 | ElevenLabs Scribe v2 | file API | 14.05 | 7.21 | 12.75 |
| 5 | Deepgram nova-3 | file API | 15.79 | 9.19 | 14.43 |
| 6 | Speechmatics enhanced | batch job | 21.06 | 9.73 | 20.29 |

For per-source metrics, use the `scores` config or see `artifacts/reports/provider_benchmark_results.md`.

## Provider Runs

| Provider | Model | Language hint | Mode | Notes |
| --- | --- | --- | --- | --- |
| Tarteel | official | `ar-SA` | websocket realtime | Official app protocol, 16 kHz PCM streaming |
| Google | Chirp 3 | `ar-XA` | Speech-to-Text v2 sync recognize | Location `us`; sync API has a 60-second input limit |
| Google | Chirp 3 | `ar-XA` | Speech-to-Text v2 StreamingRecognize | Realtime audio pacing; no blanks in clean hypotheses |
| ElevenLabs | Scribe v2 | `ar` | speech-to-text file API | Full 600-clip run |
| Deepgram | nova-3 | `ar` | listen file API | Full 600-clip run |
| Speechmatics | enhanced | `ar` | batch job API | Full 600-clip run |

No Quran-specific phrase boosting, target ayah hints, or contextual biasing were used.

## Re-Scoring

The included scorer does not require audio. It scores hypothesis files against `benchmark.jsonl` references.

```bash
cd artifacts/source_benchmark
python3 score.py --hyps ../outputs/hypotheses/hyps.jsonl
```

Primary hypothesis files:

| Provider / model | File |
| --- | --- |
| Tarteel official | `artifacts/outputs/hypotheses/hyps.jsonl` |
| Google Chirp 3 sync | `artifacts/outputs/hypotheses/google_chirp3_hyps.clean.jsonl` |
| Google Chirp 3 realtime | `artifacts/outputs/hypotheses/google_chirp3_realtime_hyps.clean.jsonl` |
| ElevenLabs Scribe v2 | `artifacts/outputs/hypotheses/elevenlabs_scribe_v2_hyps.jsonl` |
| Deepgram nova-3 | `artifacts/outputs/hypotheses/deepgram_nova3_hyps.jsonl` |
| Speechmatics enhanced | `artifacts/outputs/hypotheses/speechmatics_enhanced_hyps.jsonl` |

The normalized `hypotheses` config keeps exactly 600 rows per provider/model run. Empty provider outputs are retained as rows with `has_hypothesis=false`.

## Tarteel Streaming Probe Notes

The Tarteel websocket exposed a fixed observable `400 ms` server processing/update grid during these runs. Active probes found stable `audioProcessedMs` deltas of `400 ms` across tested client packet sizes. This is an observable server update cadence, not proof of the internal model architecture or attention chunk.

See `artifacts/reports/limit_test_report.md`, `artifacts/outputs/probes/`, and `artifacts/outputs/scores/stream_timing.txt` for details.

## Reproducibility Notes

The scripts in `artifacts/scripts/` are copied byte-for-byte from the run directory. They are intentionally not reformatted so that the uploaded harnesses remain identical to the files used for the benchmark.

To rerun provider benchmarks, obtain source audio from `Quran-Lab/quranic-asr-benchmark` and provide your own provider credentials. Credential files are intentionally excluded.

## Citation

If you use this dataset, cite this repository and the source benchmark:

```bibtex
@dataset{quran_lab_quranic_asr_provider_benchmark,
  title = {Quranic ASR Provider Benchmark Results},
  author = {Quran Lab},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Quran-Lab/quranic-asr-cloud-rawdata}
}
```