ui3TdZddlZddlZddlZddlZddlZddlZddlZdZdZ dZ e fde de de d e d e f d Zd(de d e d e d e fdZd)de de de d e fdZd*dZde de de fdZde de de fdZd+de de de de d ef dZddddddddde de d e d!e d"e d#e d$e de de de d efd%Zd&Zed'k(reyy),uDetect voice onset in a VO take and align it to the cue start. Strategy (phase 1 — anchored inside the green focal area): 1. Detect voice onset in the take file. Cascade silencedetect at a few thresholds, then fall back to an RMS envelope on decoded PCM when silencedetect doesn't find a plausible leading silence. 2. Target position inside the file = preroll_ms + anchor_offset_ms, so voice lands slightly INSIDE cue.start on the master timeline (user asked for "somewhat inside" instead of exactly at the edge). 3. Before shifting, sanity-check that the detected onset is inside the plausible green-zone window (preroll_ms ± shadow_lead on the early side, preroll_ms + cue_duration + shadow_trail on the late side). If it's outside that window, detection lied — skip the shift entirely and leave the take as recorded. Never clobber the take with a nonsense pad/trim that pushes voice outside the green zone. 4. If detection is plausible, trim (actor came in late) or pad (actor came in early) so voice lands at target_in_file_ms. Output is a re-encoded webm (libopus 128k) so edit points are sample-accurate. Requires ffmpeg on PATH. Uses numpy for the RMS fallback. N)Dg@gdddd|dd|d|dd d g }tj|d d }|j}tjd |Dcgc]}t |j d}}tjd|Dcgc]}t |j d} }|r |ddkr| syt d} t| D]A\} } | dzt|kr|| dzn| } | | z |k\s*tt| dzcSycc}wcc}w)uRun silencedetect once and return voice-onset ms if the file has leading silence starting within 10ms of t=0 AND the first non-silent burst sustains for at least min_voice_burst_s. Sub-burst transients (record-start clicks, mouse clicks captured a few ms above threshold) are skipped — the algorithm walks past to the first silence_end whose subsequent voice burst is genuinely sustained. Without this gate, a sub-millisecond click at e.g. 80 ms gets returned as voice onset and the cascade locks in a false positive (B2 from 2026-04-30). Returns -1 if no silence_end has a sustained voice burst — caller relaxes the threshold (next cascade step) or falls back to the RMS envelope detector.ffmpeg -hide_bannerz-nostats-i-afzsilencedetect=noise=z dB:duration=-fnull-T)capture_outputtextzsilence_start:\s*([\d.]+)zsilence_end:\s*([\d.]+)rg{Gz?inf) subprocessrunstderrrefinditerfloatgroup enumeratelenintround)rrrrcmdoutrmsilence_starts silence_endsINFiend next_starts 9/home/booth/atri-vox-suite/vo-booth/scripts/align_take.py_silencedetect_onset_msr,2s/ .* j %l^< O fc  C ..T =C ZZF13=Y[a1bcAeAGGAJ'cNc13=W[a1bcAeAGGAJ'cLc ~a047L ,CL)*3./!ec.6I.I^AE*s  0 0uS4Z() )* dcs !D !Dmax_duration_s window_msc ddl}d}ddddd |d |d d d dt|dddg} tj|dd}|j |j|jj|jdz }|jdk(rytdt||zdz }|j|z} | dkry|j|d| |zj| |dzj!d} tdd|z} t#|j%| d| dz} t| dzd} | | kD}|j'syt)t+|dz D]}||s ||dzs||zcSt|j-|zS#t$rYywxYw#tj $rYywxYw)uFallback onset detector — decode the first `max_duration_s` of audio to mono PCM and find the first sustained rise above the leading noise floor. Returns onset in ms, or -1 if the envelope is flat / decode failed.rNri>r r -loglevelerrorr z-tz.2fz-ac1z-arrs16lerT)rcheck)dtypeg@rr)axisgư>g@g{Gz?)numpy ImportErrorstrrrCalledProcessError frombufferstdoutint16astypefloat32sizemaxr sqrtreshapemeanrmediananyrangerargmax)rr-r.npsrr"r#pcmwinnrmshead_nfloorthreshabover(s r+_rms_envelope_onset_msrVZs  B.+w j $ sE3r7 gs  CnnSTB -- "((- 3 : :2:: F PC xx1} aR)^d*+ ,C CA1u ''3xC=((C0A5;;;C DC C9$ %F "))CL) *T 1Ed #F &LE 99;3u:> "! 8a!e y= ! u||~  **I   ( (s"FF* F'&F'*G?Gr皙?crtD]}t||t}|dkDs|cSt|}|dk\r|Sy)uReturn ms where voice starts within the file. Cascades silence thresholds, then falls back to RMS envelope. Always returns ≥ 0 (0 = file starts hot).r)_SILENCE_THRESHOLDS_DBr,_SILENCE_MIN_DURATION_SrV)rrrthronset rms_onsets r+detect_onset_msr^sJ&' C9PQ 19L 'z2IA~ c  tj|dddy#tj$r2}td|d|jj dd|d}~wwxYw)NT)r4rrzffmpeg failed (z): i)rrr= RuntimeErrorrstrip)args descriptiones r+_ffmpegrfs`^t44H  ( (^_[MQXX^^=Mds=S .+w '$,s# j 66    r_pad_mscFtdddddd|dd|d |d d d d |gdy)Nr r r0r1rjr r zadelay=|rkrlrmrnpadrprq)rrgrss r+pad_silence_at_startrwsB .+w j &* 66    r_shift_ms tolerance_msct||krtj||d}n%|dkDrt|||d}nt ||| d}t j j|t j j||||dS)u#Shift the audio timing by shift_ms (applied ABSOLUTELY to input_path). shift_ms > 0 → trim shift_ms from the start (voice plays earlier on timeline) shift_ms < 0 → pad |shift_ms| of silence at the start (voice plays later) |shift_ms| ≤ tolerance_ms → copy unchanged copyrrorv)inputoutputryactionrz)absshutilcopyfilerrrwospathabspath)rrgryrzrs r+ shift_takers 8} $ K0 A K:Zxi@,''//+.$  r_P,cue_duration_msanchor_offset_msshadow_lead_msshadow_trail_msrrrz preroll_msrrrrct|||} td||z } ||dkDr|ndz|z} | | ks| | kDrbtj||tj j |tj j |dd| | |||| | gd||d S||z} | | z }t|||| }| |d<||d<||d<||d <||d <||d <|S) a[Detect voice onset and shift so voice lands at preroll_ms + anchor_offset_ms. cue_duration_ms: width of the green focal area. When > 0, the detected onset is sanity-checked to fall within the plausible green-zone window. If it falls outside, the shift is skipped entirely (take is left as recorded). anchor_offset_ms: how far past cue.start to target voice onset. Small positive value puts voice "somewhat inside" the cue rather than exactly at the edge. shadow_lead_ms / shadow_trail_ms: tolerance on either side of the cue window for the sanity check. riskiponset_out_of_green_zone) r}r~ryrrzdetected_onset_msrrrplausible_range_msreasonrrrrrrrr)r^rDrrrrrr)rrgrrrrrrrrzonset_ms min_expected max_expectedtarget_in_file_msryinfos r+align_take_to_cuers.z<GH q*~56LOa4GTRUddL,(\"9  K0WW__Z0ggook2(!)$ 0.#/">/(*  #%55++H j+x FD (D #D/D -D 'D)D Kr_c tjtjd}|j dd|j dd|j dt d d |j d t dd |j dt dd|j dt dd|j dt dd|j dt dd|j dt dd|j dt dd |j}t|j|j|j|j|j|j|j|j |j"|j$! }t't)j*|d"#y)$Nrrpr}z Input audio file (webm/wav/etc.))helpr~zOutput webm filez --prerollTz:Expected voice-onset position in ms (the cue's preroll_ms))typerequiredrz--cue-durationzKCue duration ms (for green-zone sanity check); 0 = unknown, use wide window)rdefaultrz--anchor-offsetrz5Target voice this many ms past cue.start (default 80)z --shadow-leadrzEarly-side shadow tolerance msz--shadow-trailrzLate-side shadow tolerance msz --thresholdrz"Silence threshold dB (default -40)z --min-silencerWz$Min silence duration s (default 0.1)z --tolerancerxu-No-op if |shift| ≤ this many ms (default 5)rr7)indent)argparseArgumentParser__doc__ splitlines add_argumentr r parse_argsrr}r~preroll cue_duration anchor_offset shadow_lead shadow_trail threshold min_silence toleranceprintjsondumps)aprcresults r+_clirsw  W-?-?-A!-D EBOOG"DOEOOH#5O6OOKcDUWOO$3fhOO%CPROOO#sAaObOO$3BaObOOMuCgOhOOO%CiOjOOMQ=lOm ==?D  DKK))++''))^^&&^^ F $**VA &'r___main__)g@)rrW))rx)rrrrrrrsysrYrZ_MIN_VOICE_BURST_Sr<rr r,rVr^rfrrrwdictrrr__name__r_r+rs*  /8J%%5%QV%/4%NQ%P++s++E++TW++ad++\    5  QV  ad  ^#SsC3SCs[_>====  =  ====== =@(: zFr_