TashaSkyUp/audio-quality-dataset-nfe4-30-step2

---
pretty_name: Audio Quality Dataset NFE 4-30 Step 2
license: other
size_categories:
- 1K<n<10K
tags:
- synthetic
- audio
- spectrogram
- text-to-speech
- weak-labels
- synthetic-speech
- research
language:
- en
---

# Audio Quality Dataset: NFE 4-30 Step 2

## Overview

This dataset publishes synthetic speech artifacts and derived spectrograms used for repo-local audio-quality experiments.

At a glance:

- `2800` synthetic runs
- `200` short English prompt sentences
- `14` NFE settings: `4, 6, 8, ..., 30`
- fixed seed `1024`

Each row represents one synthetic run and includes:

- prompt text
- raw synthetic WAV
- processed synthetic WAV
- spectrogram PNG
- NFE value
- procedural weak label

Here, **NFE** means the number of LongCat inference / ODE steps used during synthesis.

## Synthetic Data Disclosure

Everything in this dataset is synthetic or derived from synthetic artifacts.

- The prompt text comes from a repo-local TSV used for this experiment.
- Raw speech was generated with `meituan-longcat/LongCat-AudioDiT-3.5B`.
- Processed speech was produced by applying one `ClearVoice` `MossFormer2_SR_48K` pass.
- Spectrogram PNGs were rendered from those processed synthetic audio files.
- No human speech recordings are distributed in this dataset bundle.
- Weak labels are procedural NFE-derived buckets, not human annotations.

## What One Row Contains

The manifest rows include:

- `run_id`
- `sentence_id`
- `category`
- `text`
- `nfe`
- `seed`
- `weak_label`
- `sentence_file`
- `run_root`
- `raw_output_dir`
- `processed_output_dir`
- `raw_wav`
- `processed_wav`
- `spectrogram_png`

Primary files in the repo:

- `manifest.tsv`
- `manifest.jsonl`
- `manifest_remote_ready.tsv`
- `summary.json`
- `sentences/`
- `runs/`

Under `runs/<run_id>/` you get the actual synthetic artifacts for that run.

## How The Dataset Was Created

This dataset was built in three stages.

### 1. Prompt and manifest build

Source prompt sheet:

- `tmp/audio_quality_dataset/short_sentences_200.tsv`

Prompt structure:

- `20` categories
- `10` sentences per category
- `200` total sentences

Manifest settings:

- NFE start: `4`
- NFE stop: `30`
- NFE step: `2`
- seed: `1024`

### 2. Synthetic audio generation

For each manifest row:

- raw synthetic speech: `LongCat-AudioDiT-3.5B`
- processed synthetic speech: one `ClearVoice` `MossFormer2_SR_48K` pass

### 3. Spectrogram rendering

Each processed synthetic clip was rendered as:

- grayscale PNG
- `1024 x 512`
- fixed dB range `[-120, 0]`
- `n_fft = 1024`
- `hop = 256`

Repo revision for this dataset export:

- `064a6bd4df88b3222459350d74341933dcfda075`

## Weak Label Semantics

The weak labels are procedural buckets derived only from NFE bands:

- `bad_like`: `nfe <= 12`
- `mid_band`: `14 <= nfe <= 22`
- `good_like`: `nfe >= 24`

Counts:

- `bad_like`: `1000`
- `mid_band`: `1000`
- `good_like`: `800`

These labels are **not** human perceptual judgments. They are synthetic proxy labels designed for downstream experiments.

## Relationship To The Published Model

This dataset was used to train the published autoencoder:

- `TashaSkyUp/audio-quality-ae-spectrogram-patches-gpu3090-best-20260409`

That model uses:

- `spectrogram_png` inputs only
- sentence-level train/validation splitting before patch extraction

It does **not** use the dataset weak labels for training.

If you build new models from this dataset, splitting by `sentence_id` is the safer default to avoid text leakage across train and validation.

## Limitations

This dataset should not be treated as:

- a corpus of real human speech
- a benchmark with human-rated quality labels
- a general-purpose speech-quality dataset

It primarily captures the behavior of one synthetic generation path:

- LongCat synthesis
- one ClearVoice post-pass
- fixed spectrogram rendering settings

## Licensing And Attribution

This dataset is marked `license: other` because it is a repo-local experiment export and does not assert a new standalone permissive license over the generated artifacts.

Synthetic generation in this workflow depended on:

- `LongCat-AudioDiT` from Meituan, specifically `meituan-longcat/LongCat-AudioDiT-3.5B`
- `ClearVoice`, specifically `MossFormer2_SR_48K`

This Hugging Face dataset repo is not an official upstream release of either dependency. Check upstream terms before redistributing or reusing generated artifacts at scale.