尚未偵測到更新。GitHub 發布版本後會顯示在這裡。
依最近已完成的 8 小時觀測回合間的 GitHub 星號增量排序
The whole Maya workflow in Blender, in one install. Free and open source, Blender 4.5 – 5.2.
翻譯準備中
Share a Claude Code session with a link. Teammates join from their own Claude Code and work in it together.
翻譯準備中
個人 GitHub 帳號下公開儲存庫的擁有者可以登入並驗證所有權，自行管理介紹，無需管理員核准。
登入並產生驗證檔案，在 30 分鐘內將其提交至此儲存庫的預設分支。其他分支或分叉儲存庫中的檔案無效。然後在下方檢查檔案，即可無需管理員核准直接管理專案。請保留檔案以維持管理權限。
如需登入，將前往登入頁面。登入後請回到此處，再次點擊所選的驗證按鈕。
在專案資料夾中執行 /nomorevibe verify，即可驗證所有權並自行管理介紹。
一分鐘完成驗證Voice activity detection for ffrwd: Silero VAD in wasm, the audio passed through untouched and every span of speech leaving as a cue.
翻譯準備中
取自儲存庫 README
正在載入完整 README。 以下為已儲存的摘錄。
Voice activity detection with Silero VAD. speech listens to an audio
stream and writes a row for every 32 ms of speech in it, so a query can
write the spans out, mark them on the clip, or hand them to a
transcriber that then spends nothing on the silence.
Requires ffrwd 0.29, whose ffrwd/wasm is 0.19.1.檢視完整 README ↗Voice activity detection with Silero VAD. speech listens to an audio
stream and writes a row for every 32 ms of speech in it, so a query can
write the spans out, mark them on the clip, or hand them to a
transcriber that then spends nothing on the silence.
Requires ffrwd 0.29, whose ffrwd/wasm is 0.19.1.
COPY (
SELECT ffrwd.merge_spans(ffrwd.vad.speech(a), max_span => 30)
FROM input('interview.mp4') f, unnest(f.audio) a
WHERE a.index = 1
) TO 'spans.ndjson'
Each row carries start_t, the second its span began, and text, the
word "speech". ffrwd.merge_spans turns the rows of a span into one
cue running from its first voiced chunk to the end of its last; written
beside the clip's video and audio, the cues land as a subtitle track
marking where anybody talks. A span longer than max_span is cut there
and goes on as the next cue.
The model scores 32 ms at a time; three parameters turn scores into
spans. threshold is the score a chunk counts as speech at, 0.5 by
default. min_silence is how much quiet closes a span, 0.1 s, so a
breath mid-sentence carries across. min_speech is the shortest span
kept, 0.25 s, so a door or a drum hit never becomes a row.
A row leaves once it is sure: once its span is long enough to keep, and
for a quiet chunk inside one, once the speech resumes. The node declares
how long that can take, 0.32 s at the defaults, so a reader pairing the
rows by time waits that much and no more. threshold can change while
a query runs; the two durations cannot, since they set that wait.
The weights are pinned in the manifest and land beside the module at install.
speech(a audio_stream, threshold DEFAULT 0.5, min_speech DEFAULT 0.25, min_silence DEFAULT 0.1) returns STRUCT(start_t number, text text)[]: a row a chunk while speech lasts. The sound is not handed
back; a reader takes it from the source.spans - every span of speech in a clip, as ndjson.speech-track - the clip with a subtitle track marking the speech.ffrwd run ffrwd/vad:spans -v source=interview.mp4 -v dest=spans.ndjson
cargo build --target wasm32-wasip2 --release