# VoxStudio offline transcription evidence / 离线转写证据

Tested / 测试：2026-10-21, public Mac release 7.0.44 (130). By VoxStudio Team, the app developer. Device and timing scope: test-results.json.

## Sources and attribution / 来源与署名

English: LibriSpeech, Vassil Panayotov, Guoguo Chen, Daniel Povey and Sanjeev Khudanpur. https://www.openslr.org/12/ ; CC BY 4.0 https://creativecommons.org/licenses/by/4.0/ . Selected from openslr/librispeech_asr, clean/test, first rows in dataset-viewer source order, until >=135 seconds including inserted silence. See en-manifest.json for exact sample IDs, speakers, source text and WAV SHA256. Original LibriVox read speech; this is not an interview.

Mandarin: Google FLEURS, Conneau et al., https://huggingface.co/datasets/google/fleurs ; CC BY 4.0 https://creativecommons.org/licenses/by/4.0/ . cmn_hans_cn/test.tsv and audio/test.tar.gz downloaded 2026-10-21. First TSV rows until >=135 seconds including silence; repeated sentences are intentionally retained. Exact filenames, source text and SHA256: zh-manifest.json. Historical sample text is quoted corpus material, not current news reporting.

Changes to both corpora: concatenate utterances in source order with one second silence, convert to mono 16 kHz PCM16 WAV. No synthetic voices. No background music or noise added. Reference files contain corpus text, not claims of perfect verbatim audio alignment; parenthetical spellings may not be spoken. Selective review notes identify real corrections and remaining errors.

## Network isolation / 断网方法

Only the official test process was launched with macOS sandbox-exec and `(version 1)(allow default)(deny network-outbound)`. Other apps and the computer network stayed online. Models were prepared beforehand. Separate experiment-mode runs measured the local speech pipeline; GUI runs checked the workflow. Automatic summary calls failed with Operation not permitted, while transcripts remained usable. That failure is retained, not removed from the screenshot.

Reproduce on a Mac with this release and prepared models:

```sh
sandbox-exec -p '(version 1)(allow default)(deny network-outbound)' /path/to/VoxStudio.app/Contents/MacOS/VoxStudio --offline-transcribe-experiment --audio /path/to/en-sample.wav --output-dir /path/to/results --modes baseline,baseline,baseline,baseline --language en
```

For Mandarin use zh-sample.wav and --language zh. The built-in experiment entry point is for reproducing this measurement; normal users should follow the article's GUI workflow. Do not replace an existing session while testing your own private recordings.

## Deliverables and limits / 文件与限制

*-original.txt and *-pipeline-result.json are unedited pipeline output. *-ui-original.txt comes from the GUI before edits. *-ui-reviewed.txt records selective edits, not a fully proofread document. *-ui-reviewed.srt is a GUI subtitle export; subtitle cues and transcript text are independent and may retain errors. *-reference.txt is corpus reference text. See revision notes before reuse.

Screenshots show only test material. The walkthrough is explicitly labelled as a screenshot walkthrough, not continuous screen recording or a speed demonstration. No elapsed time should be inferred from its slide duration.

No search-console data, rankings, conversion rates or comparative accuracy have been invented. This evidence proves the specified files and observed operations on one machine, not every language, account or configuration.

## Rerun without subtitle segmentation / 未分段重跑 (added after publication of the first draft)

zh-rerun-no-segmentation.txt and .srt: the same zh-sample.wav transcribed again on the same Mac and app build through the app's local MCP interface (language zh, speakers auto, no subtitle segmentation, no translation). The MCP export contains no speaker labels. Nine text blocks cover eleven source clips; one cue lasts 21.3 seconds. Transcript errors (讨探讨, 贸易的和) recur, so they are properties of this model on this file, not a one-off.

## English rerun with speaker labels / 英文说话人重跑

en-rerun-labelled.txt, en-rerun-no-segmentation.srt and en-rerun-segmented.srt: en-sample.wav transcribed again in the GUI (English, speaker count Auto-detect, Keep speakers anonymous, Segment subtitles off, translation off). The single label Speaker 1 was renamed to "Narrator" before the TXT export. The transcript SRT has 3 cues of 45.1–48.3 s; after More → Segment subtitles, Subtitles → SRT has 38 cues of 0.7–5.7 s. All 16 source clips are LibriSpeech speaker 6930, so one label matches ground truth; this is not a multi-speaker accuracy test.

## English-only media for the English guide

Screenshots and walkthrough videos are no longer stored in this folder: they are WebP/MP4 files on the R2 CDN (static/img/blog/20261021/, static/media/blog/20261021/). The English walkthrough is a 72 s screenshot walkthrough built from English-interface screenshots (not continuous capture); caption-preview-en.html is the English caption check page, and the Mandarin guide keeps caption-preview.html. The summary-failure image is the Summary panel cropped from the offline run's screenshot; the edit-mode image is from an English session.