---
license: apache-2.0
title: Persian ASR Triple Threat
emoji: 🎙️
colorFrom: red
colorTo: pink
sdk: gradio
sdk_version: 5.22.0
app_file: app.py
python_version: 3.11
pinned: false
---
# Persian ASR Triple Threat Leaderboard

A public leaderboard for comparing Persian speech recognition models on two complementary eval sets.

Benchmarks:

- `VisualEars6669`: 6,669 challenging Persian audio clips (10.49h) curated from the VisualEars Golden6669 benchmark, with clean, farfield, and obstructed conditions.
- `FLEURS-fa Full`: the full 4,341-row Persian FLEURS benchmark at `Reza2kn/fleurs-fa-benchmark`, scored against `raw_transcription`.

Metrics:

- VisualEars6669 fair WER/CER ignores punctuation, spaces, half-spaces, and diacritics.
- FLEURS-fa Full replaces the older 871-row FLEURS test-only score; legacy short-FLEURS numbers are intentionally removed from the page.
- Triple Threat is `60% S³ + 20% WER + 20% CER`. Each metric pillar is averaged 50/50 across VisualEars6669 and FLEURS-fa Full, preserving equal weight for both dataset splits.
- Lower is better.

Rows without a VisualEars6669 score have not yet been rerun on the expanded 6,669-row benchmark; their older VisualEars269 numbers are retained in `results.json` for provenance but are not displayed in the current leaderboard.