Skip to content

sglang-explainer-video: approve fastokens, rewords from the PR 49570 / 48928 videos - #146

Merged
BBuf merged 1 commit into
mainfrom
sev-pron-pr49570-48928
Nov 14, 2026
Merged

BBuf merged 1 commit into
mainfrom
sev-pron-pr49570-48928

Conversation

@BBuf

@BBuf BBuf commented Nov 14, 2026

Copy link
Copy Markdown
Owner

Pronunciation findings from two short explainer videos (SGLang PR 49570, the bright-image edit fix, and PR 48928, segmented fastokens for DeepSeek recipe prompts), measured 2026-11-14 on zh-CN-XiaoxiaoNeural with edge-tts and mlx-whisper. Both takes went through pron_check.py --asr, seg_check.py, line_asr.py, hang_take.py and tone_take.py.

Lexicon (pronunciation.json)

  • fastokens joins approved. The voice reads it "fast tokens", which is the reading meant; pron_check reported FAIL only because Whisper writes "Fast Tokens" with a doubled t. A respelling to "fast tokens" failed the same string compare, so it is approved as written and pinned in tests/test_explainer_pronounce.py.

references/pronunciation.md records

  • Four terms with no passing respelling, each fixed by rewording the narration while the term stays on screen:
Term Heard Reword
config (pipeline config) 看Fake, same as the PR 28185 row pipeline 配置 / 配置
thinking finting 思考模式
segments (encode segments) Sediments (OK~) 分段编码接口
PIL (PIL 图片) Pill in the carrier, Pure in the line 还是图片对象的输入
  • Six seg_check cuts and their rewords: 同一 | 步, 做归 | 一化 and 上归 | 一化, 像 | 素算 / 像 | 素值, 写 | 明了 and 推 | 出来, 从 | 头切, 字 | 符数 | 算.
  • Two line_asr mishearings, both reworded: 首字延迟 was heard as 首次延迟 (now 首 token 延迟), and 亮图必中 as 样图必重 (now 只要图够亮就会出错).

Checks

  • pytest tests/test_explainer_pronounce.py tests/test_explainer_seg_check.py tests/test_explainer_publish_check.py tests/test_explainer_tone_take.py: 47 passed before trimming the new test to one assertion; tests/test_explainer_pronounce.py alone: 35 passed.
  • pre-commit run --files on the three changed files passes.

🤖 Generated with Claude Code

…/ 48928 videos

Pronunciation findings from the PR 49570 and PR 48928 explainer videos
(2026-11-14, zh-CN-XiaoxiaoNeural):

- fastokens is read "fast tokens", the reading meant; pron_check fails
  only because Whisper writes the doubled t. Add it to `approved` and
  pin it in the tests.
- config, thinking, segments and PIL had no passing respelling; the
  narration was reworded and the terms stay on screen.
- Six seg_check cuts and two line_asr mishearings, each fixed by a reword.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@BBuf
BBuf merged commit 974861d into main Nov 14, 2026
1 check passed
@BBuf
BBuf deleted the sev-pron-pr49570-48928 branch November 14, 2026 15:32
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant