jzmzhong/segmental_accent_annotation

Segmental Accent Evaluation Dataset

This repository contains the dataset collected for the paper “Beyond Utterance-Level Speaker Similarity: Perception-Aligned Segmental Evaluation of Accent in Speech Generation”.

The dataset includes human speech recordings, speech generated with F5-TTS, MaskGCT, and SPARC, utterance transcripts, and human ratings of speaker and accent similarity. Segmental annotations identify the subword segments that listeners perceived as cues to accent differences between a reference recording and generated speech.

Explore the interactive results viewer to listen to speech pairs and view their ratings and segmental annotations.

Dataset overview

Component	Contents
Accent regions	Australian English and Scottish English
Recorded speakers	10: five Australian and five Scottish
Original recordings	340 WAV files: 34 utterances per speaker
Generated speech	1,020 WAV files: 34 utterances per speaker for each of three systems
Audio format	Mono, 16 kHz, 16-bit PCM WAV
Text	34 original transcripts and their segmented versions
Evaluated speech	200 generated utterances from four target speakers
Human responses	2,500 retained rater–utterance responses, including 2,350 stranger-voice responses and 150 voice-donor responses

The audio collection is larger than the human-evaluated subset: not every generated WAV has a corresponding annotation CSV. Each evaluated target speaker has 50 evaluated generations: 20 from F5-TTS, 20 from MaskGCT, and 10 from SPARC.

Directory structure

.
├── README.md
├── wav/
│   ├── australian/
│   │   └── {speaker}/
│   │       ├── {speaker}_{utt_id}.wav
│   │       └── cloned_{speaker}_{utt_id}_ref_{prompt_utt_id}_{model}.wav
│   └── scottish/
│       └── {speaker}/
│           ├── {speaker}_{utt_id}.wav
│           └── cloned_{speaker}_{utt_id}_ref_{prompt_utt_id}_{model}.wav
├── txt/
│   └── {utt_id}.txt
├── utterances_segmented.json
└── annotations_ratings/
    ├── australian/
    │   └── {candidate_wav_stem}.csv
    ├── scottish/
    │   └── {candidate_wav_stem}.csv
    └── summary.csv

1. Original recordings

Original human recordings are stored at:

wav/{accent_region}/{speaker}/{speaker}_{utt_id}.wav

• accent_region: australian or scottish.
• speaker: the target speaker's identifier.
• utt_id: the utterance identifier, such as comma_001 or wolf_001.

2. Generated speech

Generated speech is stored at:

wav/{accent_region}/{speaker}/cloned_{speaker}_{utt_id}_ref_{prompt_utt_id}_{model}.wav

• speaker: the target speaker whose voice is being generated.
• utt_id: the utterance spoken in the generated audio.
• prompt_utt_id: the original utterance used as the synthesis prompt.
• model: f5tts, maskgct, or sparc.

The synthesis prompt is the recording {speaker}_{prompt_utt_id}.wav in the same speaker directory. It is distinct from the listening-test comparison reference, which is identified by the reference_speech column in the annotation CSV.

3. Text and segmentation

The txt/ directory contains the original transcript of each utterance as {utt_id}.txt.

utterances_segmented.json is a JSON list with two fields per entry:

Field	Description
id	Utterance identifier matching the transcript and audio filenames.
text	Transcript with | separating the segments presented during annotation.

For example:

{
  "id": "comma_001",
  "text": "Well|, |here's| |a| |sto|ry| |for| |you|:"
}

Annotation indices refer to the zero-based positions in text.split("|"). Preserve every segment, including spaces and punctuation, when resolving indices. Removing these segments before indexing changes the mapping. Joining the segments with an empty string reconstructs the unsegmented text.

4. Ratings and segmental annotations

annotations_ratings/ contains one CSV per evaluated candidate WAV, grouped by accent region. Each CSV has the same filename stem as its candidate audio and contains one row per retained rater.

Column	Description
prolific_pid	Rater identifier.
TAR_SPK	Target speaker identifier; matches the speaker's audio directory.
rater_role	listener for a stranger-voice evaluation, or voice_donor for a speaker evaluating their own voice.
reference_speech	Original recording used as the comparison reference in the listening test.
candidate_speech	Generated WAV evaluated by the rater.
annotation	Selected segment indices separated by |, such as 0|2|9|24. An empty field means no segments were selected.
annotated_subword_count	Number of selected subword segments in that response; zero for an empty selection.
speaker_smos	Individual speaker-similarity score: 1 = completely different speakers, 5 = identical speakers.
accent_smos	Individual accent-similarity score: 1 = completely different accents, 5 = identical accents.

The two score columns contain individual responses, not means across raters. Higher values indicate greater perceived similarity. Both WAV filenames can be resolved relative to wav/{accent_region}/{TAR_SPK}/.

Annotations mark subword segments that helped the rater identify accent differences between the two recordings. They are text-segment indices, not timestamps or phoneme labels. Raters could select up to 10 segments, so these annotations represent selected perceptual cues rather than an exhaustive inventory of differences.

The per-WAV CSVs contain responses retained after the target-speaker/rater exclusions used for this release. The exclusions apply equally to annotations and scores. Attention checks and demonstration trials are not included. To select only stranger-voice evaluations, retain rows with rater_role == "listener" and prolific_pid != TAR_SPK.

Aggregate results

annotations_ratings/summary.csv reports results for each evaluated target speaker and system:

• Mean number of annotated subwords per rating.
• Mean accent-similarity score.
• Mean speaker-similarity score.

The speaker and target_id columns identify each target. Metric columns end in _f5tts, _maskgct, or _sparc.

Summary values use mean ± 95% confidence-interval half-width, calculated as 1.96 × sample standard deviation / sqrt(n), where n is the number of retained ratings for that target speaker, system, and metric. Only stranger-voice evaluations are included in the summary; voice-donor evaluations are excluded. These intervals are calculated across individual ratings and are not adjusted for repeated ratings by the same listener or of the same utterance.

License

This dataset is licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

You may share and adapt the data for noncommercial purposes, subject to the license terms. Give appropriate credit, provide a link to the license, and indicate any changes you make. See the CC BY-NC 4.0 license for the full terms.