Datasets:
File size: 8,442 Bytes
5ad61e0 78f03cc 5ad61e0 93dd4e9 bbe9e99 5ad61e0 bf21810 5ad61e0 bf21810 78f03cc bf21810 887e6ad bf21810 5ad61e0 64c694b 5ad61e0 64c694b 5ad61e0 78f03cc 5ad61e0 bbe9e99 5ad61e0 bf21810 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 | ---
license: other
task_categories:
- audio-classification
tags:
- audio
- multi-label-classification
- speech-quality
---
# Speech Quality Test Labels
PICSAFEv1 is a multi-source annotated test dataset for evaluating speech quality assessment and audio data filtering methods. It contains 10,728 audio samples drawn from 14 source datasets, with annotations from a vocabulary of 33 tags. These tags describe recording provenance, speech styles, speaking rate and pitch, speaker attributes, noise, reverberation, distortion, and transcript errors. These annotations support benchmarking quality metrics and filtering methods, selecting thresholds, analyzing errors across audio conditions, and studying data selection for automatic speech recognition and text-to-speech using independently acquired source audio.
The dataset provides a shared evaluation set for studying which audio samples a filtering system accepts or rejects, and why. Its tag annotations support both detailed analysis of individual audio characteristics and binary evaluation under the `strict`, `mid`, and `loose` filtering policies described below. Acceptance is policy-dependent: a negative label indicates exclusion under the selected policy, rather than that the recording is unsuitable for every task.
When using PICSAFEv1 for threshold selection or model development, keep a separate held-out subset for final evaluation to avoid reporting performance on the same samples used for tuning.
It publishes this README, [`metadata.jsonl`](./metadata.jsonl), and a tag-to-binary conversion script, because the 14 source datasets have different copyright and access terms.
The manifest contains the 10,728-sample PICSAFEv1 subset. It is not a license to redistribute the underlying recordings.
## Metadata fields
- `file_name`: audio basename.
- `id`: globally unique sample identifier with the `[source]` prefix.
- `source`: source dataset name.
- `source_path`: path relative to the origin directory.
- `tags`: list of labels.
## Source datasets and official download paths
| Source dataset | Samples | License / access note | Official download or access path |
| --- | ---: | --- | --- |
| CSTR-NAM-TIMIT-Plus | 842 | ODC-By 1.0 | [Edinburgh DataShare](https://datashare.ed.ac.uk/items/30d10e2e-0ccb-483b-b499-8b3c98ad0a2c/full) |
| DNS5_LibriVox | 510 | LibriVox public domain; jurisdiction restrictions may apply | [Microsoft DNS5 download instructions](https://github.com/microsoft/DNS-Challenge/blob/master/README.md) |
| EARS | 1,042 | CC BY-NC 4.0; non-commercial use | [Official GitHub repository](https://github.com/facebookresearch/ears_dataset) |
| MSceneSpeech | 1,000 | Dataset audio terms are not stated separately; confirm with authors | [Project page](https://speechai-demo.github.io/MSceneSpeech/) / [Google Drive](https://drive.google.com/drive/folders/12DSuH8ynBc82RYYHQZmMmp7q6O1xSVh5?usp=sharing) |
| NISQA (NISQA_TEST_FOR) | 724 | Original source terms; this subset is restricted to non-commercial research/forensic use | [NISQA Corpus wiki](https://github.com/gabrielmittag/NISQA/wiki/NISQA-Corpus) / [DepositOnce archive](https://depositonce.tu-berlin.de/items/b8908103-b0e8-4912-8144-aea65098fa1f) |
| SOMOSv2 | 500 | Research and non-commercial use only; redistribution under the same terms | [Zenodo record](https://zenodo.org/records/7378801) |
| TencentCorpus | 1,010 | No separate public audio license located; challenge access terms apply | [ConferencingSpeech2022 repository](https://github.com/runngezhang/ConferencingSpeech2022) / [challenge plan](https://tea-lab.tencent.com/ConferencingSpeech_2022_Challenge_Evaluation_Plan_version2.pdf) |
| VCTK | 1,000 | CC BY 4.0 for the official VCTK release; verify the exact version | [Edinburgh DataShare](https://datashare.ed.ac.uk/items/30e7453c-9ea8-48b4-8e18-f96d0dc62928/full) |
| WenetSpeech | 600 | CC BY 4.0; access requires the official form/password workflow | [OpenSLR SLR121](https://www.openslr.org/121/) |
| Whisper40 | 500 | No explicit dataset license located in the official repository | [Official GitHub repository](https://github.com/Lijingze666/whisper40) |
| WSJ | 1,000 | LDC User Agreement; license required | [WSJ0 / LDC93S6A](https://catalog.ldc.upenn.edu/LDC93S6A) and [WSJ1 / LDC94S13A](https://catalog.ldc.upenn.edu/LDC94S13A) |
| zhvoice | 500 | Mixed upstream datasets; no unified license | [Official GitHub repository](https://github.com/fighting41love/zhvoice) (Baidu download link is on the page) |
| LibriTTS-R | 500 | CC BY 4.0 | [OpenSLR SLR141](https://www.openslr.org/141) |
| LibriTTS | 1,000 | CC BY 4.0 | [OpenSLR SLR60](https://www.openslr.org/60/) |
## Label schema
In article, it is described that the 33 labels are:
`Nspk`, `distortion`, `emotional`, `enhanced`, `excessive_sibilance`, `fast_speed`, `female`, `high_pitch`, `imperceptible_noise_level`, `instantaneous_noise`, `low_noise_level`, `low_pitch`, `male`, `microphone_popping`, `mid_high_noise_level`, `music_or_effect`, `nb_noise`, `non_binary`, `non_speech`, `read_speech`, `real_recording`, `regular_pitch`, `regular_speed`, `reverberation`, `singing_voice`, `slow_speed`, `speech_overlap`, `spontaneous_speech`, `synthetic`, `vocal_sound`, `wb_noise`, `whispered_speech`, and `wrong_transcript`.
## Converting tags to binary labels
Labels are derived from the human tags, not predictions
from a decision tree or metric thresholds.
- `1`: accepted under the selected policy (positive).
- `0`: rejected because at least one negative tag is present (negative).
The label is `0` if any annotation tag belongs to the mode-specific negative
set below or the additional `--negative_tags` set; otherwise it is `1`.
| Mode | Tags that produce label `0` (any match) |
| --- | --- |
| `strict` (default) | `Nspk`, `speech_overlap`, `low_noise_level`, `mid_high_noise_level`, `reverberation`, `music_or_effect`, `microphone_popping`, `excessive_sibilance`, `distortion` |
| `mid` | `speech_overlap`, `mid_high_noise_level`, `reverberation`, `music_or_effect`, `microphone_popping`, `excessive_sibilance`, `distortion` |
| `loose` | `speech_overlap`, `mid_high_noise_level`, `reverberation`, `music_or_effect`, `distortion` |
`mid` allows `Nspk` and `low_noise_level`; `loose` additionally allows
`microphone_popping` and `excessive_sibilance`. Other tags do not independently
cause rejection. In particular, `non_speech`, `wrong_transcript`, `singing_voice`,
`vocal_sound`, `nb_noise`, `wb_noise`, and `instantaneous_noise` are **not** default
negative tags. Label `1` means policy acceptance, not a universal claim of clean
speech or a correct transcript. Add rejection tags explicitly if needed.
Names are case-sensitive (`Nspk`). An empty tag list yields `1`, matching the
reference function, but does not establish annotation completeness.
Missing tag fields and unknown tag names are errors.
Run from the repository directory:
```bash
# Default policy.
python3 tags_to_binary.py --input metadata.jsonl --output /tmp/picsafe_strict.jsonl
# Alternative policy (also supports loose).
python3 tags_to_binary.py --input metadata.jsonl --output /tmp/picsafe_mid.jsonl --positive_mode mid
# Custom policy, different from the default labels.
python3 tags_to_binary.py --input metadata.jsonl --output /tmp/picsafe_custom.jsonl --negative_tags non_speech wrong_transcript
# Original annotations: UID and semicolon-separated Tags columns.
python3 tags_to_binary.py --input annotations.tsv --output /tmp/picsafe_from_tsv.jsonl
```
Output JSONL preserves all original fields and adds `binary_label` and
`label_policy` (the selected `positive_mode` and additional `negative_tags`).
TSV input also adds `id` from `UID` and `tags` from `Tags`.
The script prints class counts, validates unique IDs and tag names, and refuses
to overwrite existing output files. The source `metadata.jsonl` is unchanged.
Join with metric scores using globally unique `id`; basenames can collide.
To reproduce a threshold-search run, use the same mode, additional negative
tags, and sample subset.
## Licensing and limitations
This repository does not grant, aggregate, or override any source-dataset license. Obtain the audio directly from the source providers and follow their current terms, attribution requirements, access restrictions, and applicable privacy/publicity rules. Some sources require non-commercial research use, registration, a license, or direct permission from the authors.
|