# VoxStudio 访谈说话人标签测试证据 / Speaker-label test evidence

测试日期 / Tested: 2026-10-21. 测试者 / By: VoxStudio Team (the app developer).

## 测试版本与设备 / App build and machine

- VoxStudio 7.0.45 (131), a pre-release build run from the project tree. It is not the public 7.0.44 (130) download used by the offline transcription guide. Re-run on the public release before publishing any version-specific claim.
- Apple M5 MacBook Pro, 32 GB memory, macOS 27.0.1 (26A434).
- The same test was also run on a 16 GB M2 Mac with macOS 27. Its exports are not in this folder yet.

## 测试素材与来源 / Test material and sources

English: `en-two-speaker-sample.wav`, 129.2 s, mono 16 kHz PCM16. Two LibriSpeech test-clean speakers, 1320 (voice A) and 1995 (voice B), with real speaker IDs. 16 read utterances, 12 alternating turns. LibriSpeech, Vassil Panayotov et al., https://www.openslr.org/12/ , CC BY 4.0 https://creativecommons.org/licenses/by/4.0/ . See `en-manifest.json` for utterance IDs, source text and turn times.

Mandarin: `zh-two-voice-sample.wav`, 173.3 s, mono 16 kHz PCM16. Google FLEURS cmn_hans_cn/test, CC BY 4.0 https://huggingface.co/datasets/google/fleurs . 16 utterances, 12 alternating turns. FLEURS does not expose speaker IDs in the distributed files used here, so voice A is the dataset's FEMALE tag and voice B is the MALE tag. This separates two different voices, not two identified people. Several people may share one tag. See `zh-manifest.json`.

Both samples are read speech, not conversations. There is no overlap, interruption, background noise or laughter. The WAV files are served from the R2 CDN (`static/media/blog/20261021/speaker-labels-*-sample.wav`), because the deploy rsync excludes `*.wav`. Their SHA256 values are in the manifests.

## 应用设置 / App settings used

Processing options: Input language set to English or 简体中文 for each file; Speaker count = Auto-detect; Recognize people = Keep speakers anonymous; Segment subtitles = off. Other speaker-count values were not tested.

## 运行记录 / Runs

| Run | Language | Output file | Blocks | Labels | Result |
| --- | --- | --- | --- | --- | --- |
| en run 1 | English | `en-run1-transcript-labelled.txt` | 12 | 2 | 16 of 16 utterances mapped to the correct voice |
| en run 2 | English | `en-run2-transcript-labelled.txt` | 12 | 2 | Text and labels identical to run 1 and run 3 |
| en run 3 | English | `en-run3-transcript-labelled.txt` | 12 | 2 | Text and labels identical to run 2; used for the screenshots |
| zh run 1 | 简体中文 | `zh-run1-transcript-labelled.txt` | 12 | 3 | Not scored for accuracy (see limits) |
| zh run 4 | 简体中文 | `zh-run4-transcript-labelled.txt` | 12 | 3 | Text and labels identical to zh run 1; used for the screenshots |

zh run 4 is the run used for the published Mandarin screenshots. Its export matches zh run 1 exactly.

## 评分方法 / Scoring method

`score_labels.py` aligns each reference utterance to the export block that covers most of its characters. It then maps labels to reference voices with the best one-to-one permutation, which is generous to the app. English: 2 labels, 2 voices, 16/16 correct, every block 100% from one voice. Mandarin: 3 labels for 2 tagged voices. Blocks 4 and 8 (1-based) contain text from both reference voices (purity 0.73 and 0.51). We do not report a Mandarin accuracy, because the reference voices are gender tags and a third label could be a real different speaker.

## 文本错误示例 / Transcript errors seen

- English: the reference spells `CHINGACHGOOK`; the export reads `Chingachguk`.
- 简体中文: 参考 `这里几乎都是沙滩` → 导出 `这里几乎都是海沙滩`；参考 `基希讷乌` → 导出 `新基纳乌`；参考 `Scotturb 403路` → 导出 `Gauterberg 403路`；参考 `但俄语也被广泛使用` → 导出 `但俄语也被广泛使。`

## 截图与视频 / Screenshots and video

Screenshots are full VoxStudio window captures (`screencapture -l`), so the computer-use status overlay is not included. The left sidebar was cropped out because it shows other session titles. The processing-options screen was not published as a still, because it displays a local file path.

`speaker-labels-screenshot-walkthrough.mp4` plays four captured screens for 5 seconds each. It is a screenshot sequence, not a continuous screen recording. Continuous recordings were made during the runs but are not published: the computer-use status panel and an unrelated chat window overlapped the VoxStudio window on several frames.

## 复现 / Reproduce

1. Use a recording you are allowed to use, or the two WAV samples from the CDN.
2. In VoxStudio, import the file, set the input language, keep Speaker count = Auto-detect and Recognize people = Keep speakers anonymous, then transcribe.
3. Export: Transcript · Original · TXT · With speaker labels.
4. Score against the manifest: `python3 score_labels.py en-manifest.json <export>.txt` (English). Mandarin scoring is not claimed.

## 限制 / Limits

- Not an interview: read speech only. Overlap, interruptions, crosstalk and noisy rooms are untested.
- Two voices only; speaker-count settings other than Auto-detect were not measured.
- Mandarin voices are gender tags, not speaker IDs.
- No timings are reported here.
- This test did not run with the network disabled. The offline transcription guide covers the isolation test.
- The AI summary panel appears automatically when configured. Its wording is generated and was not scored.