---
license: other
task_categories:
- audio-to-audio
language:
- en
tags:
- speech-restoration
- speech-enhancement
- benchmark
pretty_name: Diamond Benchmark — 750 real degraded speech clips
size_categories:
- n<1K
---

<div align="center">

# 💎 Diamond Benchmark

### 750 real degraded speech recordings for evaluating restoration models

[![Clips](https://img.shields.io/badge/clips-750-2563eb)](.)
[![Audio](https://img.shields.io/badge/audio-real_degraded-1d4ed8)](.)
[![Metrics](https://img.shields.io/badge/DNSMOS-%2B%20CER-1e40af)](.)
[![Model](https://img.shields.io/badge/🤗-Diamond-yellow)](https://huggingface.co/nineninesix/diamond-1.0)

</div>

---

These are **real, genuinely degraded** speech recordings — not synthetic, not
clean audio re-degraded by a pipeline. Each clip carries the damage of an actual
real-world capture: low-bitrate codec compression, narrow bandwidth, background
noise, and clipping. Together they form a **fixed benchmark** for measuring how
well a speech-restoration model recovers clean, studio-quality audio from real
degradation.

Every clip ships with its reference transcript, so restoration can be scored on
**two axes at once** — perceptual quality *and* content preservation.

## What's inside

| | |
|---|---|
| **Clips** | 750 degraded `.wav` recordings |
| **Language** | English |
| **Audio** | real degraded speech (codec · band-limit · noise · clipping) |
| **Reference** | ground-truth transcript per clip |

## Structure

```
diamond-bench/
├── audio/
│   └── <set>_<id>.wav      # 750 degraded input clips
└── manifest.csv            # per-clip metadata
```

`manifest.csv` columns:

| column | meaning |
|---|---|
| `set` | source subset the clip belongs to |
| `id` | clip id within its set |
| `emolia_id` | original recording id |
| `audio_path` | relative path under `audio/` |
| `sample_rate` | clip sample rate (Hz) |
| `duration_sec` | clip length (seconds) |
| `speaker` | speaker id |
| `text` | reference transcript (used for CER) |

## How to score

Two complementary metrics — one alone is not enough:

- **DNSMOS-P.835** (`sig_bak_ovr.onnx`, raw windowed mean) — perceptual quality
  (SIG / BAK / OVRL). How good the audio *sounds*.
- **CER** (ASR vs. `text`) — content preservation. A generative restorer can
  sound clean while smearing or dropping words; DNSMOS is blind to that, CER is
  not.

A genuine gain means **OVRL rises without CER rising**.

## Loading

```python
import pandas as pd
import soundfile as sf
from huggingface_hub import snapshot_download

root = snapshot_download("nineninesix/diamond-benchmark", repo_type="dataset")
meta = pd.read_csv(f"{root}/manifest.csv")

row = meta.iloc[0]
wav, sr = sf.read(f"{root}/{row.audio_path}")
print(row.id, f"{row.duration_sec:.1f}s —", row.text)
```


## Links

- **Model:** [`nineninesix/diamond-1.0`](https://huggingface.co/nineninesix/diamond-1.0)
- **Training:** [`diamond-train`](https://github.com/nineninesix-ai/diamond-train.git)
- **Inference:** [`diamond-inference`](https://github.com/nineninesix-ai/diamond-inference.git)