File size: 9,173 Bytes
da16c69 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 | # Dataset Setup
We open-source our preprocessed datasets in the S3 link here. Download it at put them into the `data` directory.
Below are the details of how we preprocessed the dataset.
## Preprocessing
### VCTK (English multi-speaker, multi-accent)
We use a curated subset of the **CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit (version 0.92)** [Yamagishi et al., 2019](https://doi.org/10.7488/ds/2645).
The original VCTK corpus contains ~44 hours of English read speech from 110 speakers with diverse accents (American, British, Irish, Scottish, Indian, etc.), each reading about 400 phonetically rich sentences recorded with high-quality microphones in a hemi-anechoic chamber.
In our benchmark, we select **40 speakers** to balance **gender** and **accent** while keeping the dataset compact:
- **Speakers:** 40 in total (22 female, 18 male)
- **Accents:** 12 accent categories
- Major accents (American, English, Canadian, Irish, Scottish, South African) have 4–6 speakers each
- Minor accents (e.g., Australian, New Zealand, Welsh, Northern Irish, Indian, British RP) have 1–3 speakers each
All audio is used at **48 kHz, 16-bit PCM, mono**, following the original VCTK speaker IDs (e.g., `p294`, `p303`, `s5`).
### LibriTTS (English multi-speaker TTS)
We use a small, speaker-balanced subset of the **LibriTTS** corpus.
LibriTTS is a large-scale, multi-speaker English corpus of read audiobooks, derived from LibriSpeech and released at 24 kHz for TTS research: it contains approximately **585 hours** of speech from **2,456 speakers**, with sentence-level segmentation, normalized text, and noisy utterances removed.
Official LibriTTS page: [https://www.openslr.org/60/](https://www.openslr.org/60/)
In our benchmark subset:
- **Speakers:** 40 total, all from LibriTTS `dev` and `test` splits
- **Gender balance (overall):**
- 20 female speakers
- 20 male speakers
- **Split design (speaker-disjoint):**
- **dev:** 19 speakers (7 F / 12 M)
- **test:** 21 speakers (13 F / 8 M)
All audio is used at **24 kHz, mono, 16-bit PCM**, following the original LibriTTS speaker IDs
### AISHELL-1
We use a subset of the **AISHELL-1** Mandarin corpus for development and evaluation.
AISHELL-1 is an open-source Mandarin speech corpus containing about **178 hours** of transcribed speech from **400 speakers**, recorded in quiet indoor environments and released at **16 kHz**. The texts cover 11 application-oriented domains (finance, science & technology, sports, entertainment, news, etc.).
Official AISHELL-1 page: https://www.openslr.org/33/
In our benchmark subset:
- **Speakers:** 40 total (IDs `S0724`–`S0763`)
- **Gender balance:** 28 female, 12 male
- **Per-speaker content:** 50 paired utterances (`gt`, `ori`) per speaker
- **Sampling rate:** 16 kHz, mono, 16-bit PCM
### UEDIN_bilingual_96kHz (Mandarin/English bilingual)
We use the **Mandarin talkers** portion of the 96 kHz release of the **EMIME Bilingual English–Mandarin Database**.
The original EMIME Mandarin/English database contains studio recordings of 14 native Mandarin speakers (7 female, 7 male) who read parallel material in both Mandarin and English for research on personalized speech-to-speech translation and cross-lingual voice conversion.
Official EMIME bilingual database page: https://www.emime.org/participate/emime-bilingual-database.html
In our benchmark subset:
- **Speakers:** 13 total from `Mandarin_talkers/`
- Female: 6 (`MF1`, `MF3`, `MF4`, `MF5`, `MF6`, `MF7`)
- Male: 7 (`MM1`–`MM7`)
- We exclude **MF2** because some of her Mandarin recordings exhibit abnormal behaviour.
- **Gender balance:** 6 female, 7 male
- **Bilingual content per speaker:**
- 25 **English → Mandarin** sentence pairs
- 25 **Mandarin → English** sentence pairs
→ 50 bilingual pairs per speaker
- **Sampling rate:** 96 kHz, mono, 16-bit PCM
### CommonVoiceFR_dev (French crowd-sourced speech)
We use a subset of **Mozilla Common Voice French v23.0**, taking clips from the `validated.tsv` split.
Official Common Voice site: [https://commonvoice.mozilla.org/](https://datacollective.mozillafoundation.org/datasets/cmflnuzw5ahjms0zbrcl0vg4e)
In our benchmark subset:
- **Source split:** `validated.tsv` from Common Voice French v23.0
- **Speakers:** 40 unique French speakers
- **Per-speaker content:** 50 utterance pairs per speaker.
- **Audio format:** mono, 16-bit PCM, 48 kHz (downsampled in code if needed)
### Long_context (LibriSpeech-Long subset)
We use a long-form English subset derived from **LibriSpeech-Long**, a benchmark dataset for long-form speech processing released with *“Long-Form Speech Generation with Spoken Language Models”* (Park et al., 2024).
- Original LibriSpeech-Long card: https://huggingface.co/datasets/ilyakam/librispeech-long
In our benchmark subset (“Long_context”):
- **Language:** English
- **Speakers:** 10
- Speaker IDs: `1272, 1673, 1919, 1993, 2078, 3576, 3853, 422, 6241, 8842`
- **Pairs per speaker:** 2 long-form utterance pairs (e.g., original vs processed)
- **Total pairs:** 20
- **Audio duration:** long-form segments, typically **0.4–4 minutes** per file
- **Sampling rate:** 16 kHz, mono, 16-bit PCM (inherited from LibriSpeech)
### VCTK natural_noise (VoiceBank+DEMAND subset)
We use a noisy VCTK subset derived from the **“Noisy speech database for training speech enhancement algorithms and TTS models”** (also known as the VoiceBank+DEMAND dataset).
The original database provides parallel **clean/noisy speech at 48 kHz** based on the VCTK multi-speaker corpus, where clean VCTK utterances are mixed with environmental noises (mainly from the DEMAND corpus) and additional speech-shaped / babble noise, and is widely used for speech enhancement and noise-robust TTS research.
Official dataset page: https://datashare.ed.ac.uk/handle/10283/2791
In our benchmark subse:
- **Language:** English (multi-accent, inherited from VCTK)
- **Speakers:** 20 VCTK speakers (`p226`–`p273`)
- **Background noise:** 10 natural noise environments, all mixed at **10 dB SNR**
- `babble`, `cafeteria`, `car`, `kitchen`, `meeting`, `metro`, `restaurant`,
`ssn` (speech-shaped noise), `station`, `traffic`
- **Per-speaker structure:**
- 20 **noisy–clean pairs** (each noisy utterance has a clean VCTK reference)
- 10 additional **clean-only** utterances
- → 30 sentence-level items per speaker
- **Overall size:** 20 speakers × 30 items ≈ **600 sentence items** (clean/noisy together give about **1,200 audio files**)
- **Audio format:** 48 kHz, mono, 16-bit PCM
### Multispeaker_libri (English multi-speaker interference)
We use a synthetic multi-speaker interference set built from our **LibriTTS** English audiobook subset.
Clean target utterances from 10 LibriTTS speakers are mixed with interfering speech from 2 additional speakers at controlled SNR levels, producing parallel **clean / mixture** pairs for studying multi-speaker robustness.
In our benchmark subset:
- **Target speakers:** 10 (5 male, 5 female)
- Male: `61, 908, 2300, 2830, 7729`
- Female: `237, 1221, 1284, 4970, 6829`
- **Interferer speakers:** 2
- Female interferer: `121`
- Male interferer: `672`
- **Clean groundtruth segments:** 100 (10 per target speaker)
- **Mixtures per groundtruth:** 8
- 4 SNR levels: `-5 dB`, `0 dB`, `+5 dB`, `+10 dB`
- 2 interferers: `121`, `672`
- **Total entries in `manifest.json`:** 800 mixture entries (each with a linked clean groundtruth)
- **Sampling rate:** 16 kHz, mono, 16-bit PCM
### iemocap (TO BE EXPLORED — emotional English speech)
> Status: **to be explored** — this dataset is prepared but not yet integrated into the main benchmark pipeline.
We construct an emotional subset from the **IEMOCAP (Interactive Emotional Dyadic Motion Capture)** corpus.
In our prepared subset:
- **Language:** English
- **Speakers:** 10 total (5 male, 5 female)
- Female: `Ses01_F, Ses02_F, Ses03_F, Ses04_F, Ses05_F`
- Male: `Ses01_M, Ses02_M, Ses03_M, Ses04_M, Ses05_M`
- **Total pairs:** 184 (`ori`, `gt`) audio pairs
- **Emotion categories (6):** `ang`, `exc`, `fru`, `hap`, `neu`, `sad`
**Pair construction rule**
For each speaker, we select utterances from the six emotion types and build pairs such that:
- `ori` and `gt` always come from the **same speaker**, but
- **their emotions are deliberately different** (e.g., `ori = ang`, `gt = neu`).
### robocall_ftc (TO BE DONE — real-world scam calls)
> Status: **to be done** — planned for trustworthy deepfake / telecom-fraud experiments.
We plan to integrate the **Robocall Audio Dataset** released by Prasad & Reaves (NCSU) based on the FTC’s *Project Point of No Entry* (PPoNE).
This dataset contains real-world audio recordings of automated or semi-automated phone calls (“robocalls”), most of which are suspected to be **illegal scam or spam calls**.
Official repo: https://github.com/wspr-ncsu/robocall-audio-dataset
Planned usage in our benchmark:
|