#iJ!dZddlZddlZdZdZdZdZdZdZdZ dZ dZ dZ dZ d Zd Zd Zd Zd ZdZdZdZeeeeeee e e eedf dedededededededededededededed ed!ef d"Zd#Zed$kr edSdS)%u)Auto-correct cue start / end timecodes by scanning a reference waveform. Primary algorithm — **hysteresis edge detection** (requires a clean voice-only stem, e.g. HTDemucs `voice_waveform.json`): - Cue START = silence → voice rising edge closest to the claimed start. - Cue END = voice → silence falling edge closest to the claimed end. Thresholds are derived dynamically from the search window's peak distribution so the detector auto-scales across quiet dialogue and shouted scenes. Fallback — **nearest local minimum** (for intra-phrase cue splits where the voice is continuous across the bound, so no rising/falling edge exists): - Smooth the peaks (box filter), enumerate local minima in the window, rank by depth + proximity to the claimed bound, return the best one. The fallback is what v1 used exclusively when the waveform was mixed audio and music made edge detection unreliable. Now that we read voice-only stems, edges are usable and preferred. Waveform structure: { "sampleRate": 100, "peaks": [0..1, ...] } at 10 ms steps. Nii xig333333?(g?c tdtt|dz |z}|dz}t|}dg|z}t |dt ||dz}t ||dz}t |D]C} ||z || <| |zdz} | |z } | |kr||| z }|dz }| dkr||| z}|dz}D|S)Nr)maxintroundlensumminrange) peakssr smoothing_msswhalfNout window_sumcountiadd_irem_is [/Users/a11/Documents/CLAUDE_ENVIRONMENTS/ATRI-VOX-SUITE/vo-booth/scripts/autocorrect_cue.py _smoothedr6s QE,-23344 5 5B 7D E A %!)CU1SD1H---.//J 4!8  E 1XX33e#AD1 D 199 %, &J  A:: %, &J  Jc ^g}d}tt|dzdt|t|dz D]h}||||dz krQ||||dzkr<|#||z |kr||||kr||d<|}Q|||}i|S)Nr)rr rrappend)smoothi0i1 min_spacingminimaprevks r _local_minimar+Hs F D 3rAvq>>3r3v;;?#;#; < <'' !9va!e} $ $fQUm)C)CQX$<$<!9vd|++!"F2J1D MM!   Qd Mr c tt||z}tt||z} td|| z } tt ||| z} | | dzkrdSt |||t| | } | dkrdStdtt|dz |z} t | | | }| |zfd|D}|sdStdtt|dz |z}d}td}|D]1}t||z }||z | z}||z}||kr|}|}2|S)Nrrrc,g|]}|k|Sr/).0r* depth_capr$s r z%_find_nearest_min..es'BBB6!9 +A+A!+A+A+Ar inf) r r r rrrr+floatabs)rrtarget_swindow_srmin_spacing_ms depth_factordistance_bias_mstarget_iwinr%r& window_maxr' candidates bias_samplesbest best_scorer*dist dist_penaltyscorer1r$s @@r_find_nearest_minrEUs5B''((H eHrM"" # #C Q3  B SZZC ( (B R!V||t ub, / /FVBrE]##JQtaU>D#82#=>>??@@Kvr2{;;J\)IBBBBBZBBBJ tq#e$4t$;b$@AABBCCL DuJ ))1x<  |+z9 q L( :  Jq Kr c *|||}|sdSt|}|tt|dz tt|dz}t d|t z}t d|t z}||fS)N)g{Gz?gQ?rg?g{Gz?g{Gz?)sortedrrr r DEFAULT_VOICE_FACTORDEFAULT_SILENCE_FACTOR)rr%r&windowsorted_wp90 voice_threshsilence_threshs r_dynamic_thresholdsrOxs 2b5\F zf~~H 3s8}}q(#c(mmc.A*B*BCC DCtS#7788L|.DDEEN  ''r c >tt||z}td|tt||zz }tt ||tt||zz} | |z dkrdSt ||| \} } tdtt|dz |z} tdtt|dz |z} d}t d}t|| z| }t| | z t || z }t||D]}d}t||| zD]}||| kr|dz }|| dzkr5d}t|| z |D]}||| kr|dz }|| dzkrht||z }||kr|}|}|dS|}|dkr)||dz | kr|dz}|dkr||dz | k|S)uFind silence → voice transition index closest to target_s within the asymmetric window [target - back_s, target + forward_s]. Two-pass detection: 1. Anchor scan — find a position where sustained voice follows sustained silence (robust against mid-phrase peaks / HTDemucs bleed). 2. Onset snap — from the anchor, walk BACKWARD through every non-silent sample to find the true first audible sample of that voice burst. This removes the "lands late" bias: if a phrase ramps up gradually, the anchor sits where ≥70% of the next 120 ms is above voice_thresh, which is already ~60 ms into the speech. Walking back through the ramp snaps to the earliest non-silent sample without any directional preference. rNrrr3gffffff? r r r rrrOr4rr5)rrr6back_s forward_s voice_min_mssilence_min_msr;r%r&rMrN voice_min silence_minr@ best_distlohir voice_countj silence_countrBonsets r_find_rising_edger`sb 5B''((H Q3uVb[11222 3 3B SZZCi"n(=(=$>$>> ? ?B Bw||t#6ub"#E#E L.As5!4r!9::;;<D#82#=>>??@@K De I R+ { + +B R)^SZZ)3 4 4B 2r]]'' q!i-(( ! !AQx,&&q S ( (  q;** # #AQx.((" ;, , , 1x<   )  IQd |t E !))eai(>99   !))eai(>99 Lr c tt||z}td|tt||zz }tt ||tt||zz} | |z dkrdSt ||| \} } tdtt|dz |z} tdtt|dz |z} d}t d}t|| }t| | z t || z }t||D]}d}t||| zD]}||| kr|dz }|| dzkr5d}ttd|| z |D]}|||kr||}|| krvt||z }||kr|}|}|dSt |}|}||kr#||| kr|dz }||kr ||| ktd|| z }||kr)||dz | kr|dz}||kr||dz | k|S) uFind where voice gives way to sustained silence, closest to target_s. Mirror of `_find_rising_edge` for the trailing boundary: 1. Anchor scan — find a position where sustained silence follows recent voice activity. Voice endings TAPER (loud → quiet → silent over ~500 ms), so there's no sharp edge. Anchor detects: (a) sustained silence of `silence_min_ms` starting at index i, and (b) voice activity somewhere in the preceding `voice_lookback_ms` window. Max-based lookback tolerates taper — the phrase can be quiet at i-1 as long as it was loud earlier in the lookback. 2. Offset snap — from the anchor, refine so the returned index is exactly "one past the last audible sample". The 85%-silent rule in the anchor tolerates a few stray tail peaks inside [i, i+silence_min], and the anchor may land a few samples after the true tail because the silence needed to sustain. Snap: (a) walk FORWARD while peaks[i] > silence_thresh — captures any straggler tail peaks the 85% rule allowed past the anchor; (b) walk BACKWARD while peaks[i-1] ≤ silence_thresh — pulls back past any silent samples inside the anchor region so the result lands immediately after the last audible peak. The back-walk is bounded to the lookback window so it can't cross a real silence gap into a previous voice burst. rrQNrrr3g333333?r rR)rrr6rSrTrVvoice_lookback_msr;r%r&rMrNrXvoice_lookbackr@rYrZr[rr^r]prev_maxrBroffset back_limits r_find_voice_endrgs65B''((H Q3uVb[11222 3 3B SZZCi"n(=(=$>$>> ? ?B Bw||t#6ub"#E#E L.aU>D#82#=>>??@@KC&7$&>&C D DEEFFN De I R B R+ s5zzK7 8 8B 2r]]'' q!k/** # #AQx.((" ;- - - s1a.011155 $ $AQx("" 8 l " " 1x<   )  IQd |t E A F 1**v77!  1**v77 Q~-..J :  % "3~"E"E!  :  % "3~"E"E Mr r waveformoriginal_start_soriginal_end_s start_back_msstart_forward_ms end_back_msend_forward_msrUrVrbrr8r9r:gate_thresholdreturncDt|dpd}|d}rdkrfd|D}|dz }|dz }|dz }|dz }||zdz }t|||||||}|dnd}|t||||| | | | }|d nd }t ||||||| }|d nd}|t||||| | | | }|d nd }|||z nd}|||z nd}||n|}||n|}||krKt ||z }t ||z }||kr |}|r|d nd }n |}|r|d nd }||kr|}|}||||tt ||z d ztt ||z d z|du|du||d||||||| | | | | d d S)N sampleRatedrrc$g|] }|kr|nd S)r r/)r0pros rr2z#autocorrect_cue..(s'BBBqa>))sBBBr g@@g@ rising_edge local_minnone voice_end _revertedr)startend rkrlrmrnrUrVrbrr8r9r:ro) original_start original_endcorrected_start corrected_enddelta_start_ms delta_end_ms onset_found offset_foundmethodparams)r getr`rErgr5r ) rhrirjrkrlrmrnrUrVrbrr8r9r:rorr start_back_sstart_forward_s end_back_s end_forward_sfallback_window_s start_idx start_methodend_idx end_methodrr final_start final_end start_delta end_deltas ` rautocorrect_cuers X\\, ' ' .3 / /B W E C.1,,BBBBEBBB#/L&/O!/J$/M%73>"%-=".".@@I%.$9==tL%eR1ACT&2N&24DFF '0&;{{ eR(-,.?AAG!( 3J#E2~?P$0.$02BDD%,$7[[V *3*?y2~~TO&-&9Wr\\tM%4%@//FVK%2%@--nIK+(8899 )677  " "*K9EQl55556LL&I5?KJ1111VJ  # #*K&I+(&$e[3C%Ct$KLLMMeY>%At$KLLMM)5'5(<<* 0&,(,!2(,( 0,     r ctjtd}|d|dt d|dt d|dt t |d t t |d t t |d t t |d t t |dt t |dt t |dt t |dt t |dt t |dt t" |dt dd|}t'|j5}t+j|}dddn #1swxYwYt/||j|j|j|j|j|j|j|j|j |j!|j"|j#|j$|j%}tMt+j'|ddS)Nr) descriptionrhz--startT)typerequiredz--endz--start-back-ms)rdefaultz--start-forward-msz --end-back-msz--end-forward-msz--voice-min-msz--silence-min-msz--voice-lookback-msz--smoothing-msz--min-spacing-msz--depth-factorz--distance-bias-msz--gate-thresholdr z4pre-zero peaks below this value (0..1); 0 = disabled)rrhelpr}r )indent)(argparseArgumentParser__doc__ splitlines add_argumentr4r DEFAULT_START_BACK_MSDEFAULT_START_FORWARD_MSDEFAULT_END_BACK_MSDEFAULT_END_FORWARD_MSDEFAULT_VOICE_MIN_MSDEFAULT_SILENCE_MIN_MSDEFAULT_VOICE_LOOKBACK_MSDEFAULT_SMOOTHING_MSDEFAULT_MIN_SPACING_MSDEFAULT_DEPTH_FACTORDEFAULT_DISTANCE_BIAS_MS parse_argsopenrhjsonloadrr{r|rkrlrmrnrUrVrbrr8r9r:roprintdumps)apargsfwfresults r_clirzs  W-?-?-A-A!-D E E EBOOJOOIEDO999OOGEDO999OO%s>SOTTTOO(s>VOWWWOOOs>QORRROO&s>TOUUUOO$s>ROSSSOO&s>TOUUUOO)s>WOXXXOO$s>ROSSSOO&s>TOUUUOO$u>ROSSSOO(s>VOWWWOO&ucOQQQ ==??D dm   Yq\\  DJ(.$*&*0&*&.*F $*VA & & &'''''s>II#&I#__main__)rrrrrrrrrrrHrIrrrrrr+rErOr`rgdictr4r rr__name__r/r rrs.   $   F(((:::zMMMj*?,D':*@(<*@-F(<*@*>,D,/ccdcecUc#&c&)c"%c%( c #& c %( c(+c#&c%(c#(c'*c%*c59ccccL$($($(N zDFFFFFr