6ESj=7 dZddlmZddlZddlZddlZddlmZmZddl Z dZ dZ dZ dZd \ZZd Zd Zd Zd ZdZej,dZeGddZd\ZZd\ZZd\ZZdZdZ dZ!dddddddddZ"ddddZ#edGd d!Z$ejJZ&e$a'd)d"Z(d*d#Z)d+d$Z*d,d%Z+d,d&Z,Gd'd(Z-y)-uKVocal-tone reader: HOW Ahmed said it, not WHAT he said. The STT pipeline reduces every utterance to bare text and throws the voice away — pace, pitch, loudness, laughter, all gone before any brain sees it. This module recovers a few cheap acoustic facts from the SAME utterance buffer the speaker-lock already runs on (main.py) and turns them into a one-line descriptor like "fast, urgent" or "slow, flat" that rides along with the transcript into both brains. Design notes: - **Pure numpy, ~2 ms per utterance.** No model, no torch, no GPU, no numba JIT (librosa.yin would add a 6.5 s first-call compile — not worth it). About 7x cheaper than the TitaNet speaker embedding that already runs every turn. - **We do NOT classify emotion.** Sarcasm/joking aren't decidable from sound alone — the descriptor reports the delivery (e.g. "slow, flat") and Claude infers intent from delivery + words + context, the way a person does. - **Relative to the speaker.** Everyone sits at a different pitch/pace/volume, so we track a rolling EMA baseline of Ahmed's own NEUTRAL delivery and describe deviations from it. The baseline only learns from neutral, clean, loud-enough turns so one shouty or one whisper-quiet utterance can't drag it. - **Pitch uses a periodicity gate.** A frame only contributes an F0 if its autocorrelation peak is a strong fraction of zero-lag (real voicing); noise — including a quiet clip that main.py amplified ~10x — has a weak peak and is rejected, so amplified hiss can't fake a huge pitch range. Pitch is only trusted when enough voiced frames survive; otherwise pitch tags are skipped. - **Rate & pitch dynamics are the reliable dimensions** (the AEC mic's AGC flattens absolute volume, so loudness is a soft, baseline-relative hint). - **Never throws into the run loop.** Any failure returns a neutral read. ) annotationsN) dataclassfieldi>ig{Gz?gffffff?)KiJg333333?g?g{Gz?z3(?:ha){2,}|(?:he){2,}|\bl(?:ol|mao|mfao)\b|\bhaha\bcXeZdZUdZdZded<eeZded<ee Z ded <y) ProsodyuOne utterance's read. `note` is None when delivery is unremarkable — the caller injects nothing on neutral turns so the brain isn't spammed.Nz str | Nonenote)default_factoryz list[str]tagsdictfeats) __name__ __module__ __qualname____doc__r __annotations__rlistr rr;/Users/ahmed/devFolder/Ultron/claude-voice/voice/prosody.pyr r 9s0OD*D1D)1-E4-rr )g?g?)皙ٿg)g?rg?皙??gffffff?g333333g)animatedloudurgentfastlaughingflatquietslow)r rr$T)frozenceZdZUdZdZded<dZded<dZded<dZded <e dd Z e dd Z e dd Z e dd Z e ddZe ddZe ddZy) ToneStateu$A smoothed read of Ahmed's recent vocal energy — the snapshot the TTS colours delivery with. `energy` < 0 = flat/tired, > 0 = animated; `pace` < 0 = slow; `laugh` decays from 1.0 the turn he laughs. Frozen so a worker thread can read it while the run loop swaps in the next one.floatenergypacelaughrintnc(|jtk\SN)r*_TONE_UPselfs rupz ToneState.upds!%!88rc(|jtk\Sr0)r* _TONE_HIGHr2s rhighzToneState.highf#';;*#<r r$r!rrrr'r'YsDFED%E5 AsJ 88 << << :: :: :: ??rr'ctS)uLatest smoothed vocal-tone snapshot (never None; neutral until warmed). O(1), lock-free read — safe from the TTS worker thread at synth time.) _TONE_STATErrr tone_staterKxs  rc L|jxsg}d|vxs$t|jjd}t dt dt d|D}t dt dt d|D}t5t}t|jt||jz zz|jt||jz zzt |rdnd |jtz|jd z ad d d y #1swYy xYw) zFold one read into the rolling tone EMA and publish the new snapshot. Side-effect only; reuses the tags read() already produced, so there's no extra DSP and no new call site anywhere.r!r,g333333g333333?c3HK|]}tj|dywr(N)_TONE_ENERGY_Wget.0ts r z _publish_tone..s#M1N$6$6q#$> "rrc3HK|]}tj|dywrN) _TONE_PACE_WrPrQs rrTz _publish_tone..s#KdL$4$4Q$)lenFRAMEnpzerosfloat32reshapesqrtmean)audior.frs r _frame_rmsr~sl E ee#AAvxx,, r  2u %B 77B!G>>q>)D0 11rc|ddtz}t|tztz}|dk(r%tjdtj S|d|j dt}ttzttz}}g}|D]}tj|dzjtkr1||jz }tj||dtdz d}|d}|d krp|||} | jdk(rt| j} | | |z t kr|j#t|| zz tj$|tj S) u?Per-voiced-frame F0 via autocorrelation with a PERIODICITY gate: a frame only counts if its best in-band autocorr peak is >= _PERIODICITY of zero-lag. That rejects noise/unvoiced frames (weak peak) — including amplified silence — so hiss can't fake pitch. Analyses up to the first 5 s for a stable read.NrrnrprqfullrXgư>) SAMPLE_RATErtrurvrwrxry_F0_HI_F0_LOrzr{ _VOICED_GATE correlatesizer-argmax _PERIODICITYappendasarray) r|r.frameslohioutfaczsegks r _pitch_trackrsC "1{? #E E ee#AAvxx,, 2AY  r5 )F F "K6$9BC  77AF==? #l 2  L \\!Q ' 3 qE 9 Bi 88q=    q6A: $  ;"q&)* ::c ,,rc:eZdZdZddZddZ d d dZd dZy) ProsodyReaderzReads vocal tone off raw utterance audio, relative to a rolling baseline of the owner's own neutral delivery. One instance per session; call read() per turn (mirrors SpeakerGate's construct-once/check-per-turn shape).cd|_d|_d|_d|_d|_t j jdddk7|_y)NrPROSODY10) _n _base_wpm _base_prange _base_loud _base_evarosenvironrPenabledr2s r__init__zProsodyReader.__init__sA'+*.(,(,zz~~i5< rc(|jdu}d}|r|n||j||_||j||_|r|n||j||_|r|n||j||_|xjdz c_y)Nc0||S||S|t||z zzSr0)_EMA)rbnews rstepz,ProsodyReader._update_baseline..steps){ +3 C3s1C+C CrrX)rrrrr)r3wpmprangerevarfirstrs r_update_baselinezProsodyReader._update_baselines|$& D!&4+D !2!2F;"'$T$//4-H"'$T$//4-H 1 rNc|js tS |j||||} t ||S#t$r tcYSwxYw#t$rY|SwxYw)zaudio: float32 mono @ 16 kHz utterance. speech_s: net voiced seconds. text: transcript. loud: PRE-normalization peak/level (main.py has it); falls back to this clip's RMS. Never raises.)rr _read Exceptionrc)r3r|speech_srkrr_s rreadzProsodyReader.readsq ||9  E8T48B  "   9      s : AAA A A c Jt|}t|xsdj}|Ttt j t j |jtjdzdz}|tks|dk(r!|rtddgddiStd|d Sd |zt|d z }t|}|jr-t|j|j dzz nd } t|} | jt k\rHt j"| d dg\} } t| | z } tt j$| }ndx} }t'|t'|d|r t'|nd| t'| ndt'| dt)| j||j*d}|t,k\xr| jt k\}|j*t.ks |j03|r|j3|| || |rdnd}t||rdg|Sg|Sg|rj5dt7|j8xr||j8dzkD}|dk\rK||j0dzkDrj5|rdndn#||j0dzkrj5d| |j:r|j<s| |j:dzkDxs| |j<dzkD}| |j:dzkxr| |j<dzk}|rj5dn|rj5d|r&t?fddDsj5dn/|j8r#||j8d zkrj5d!t?fd"d#D}|s|r|j3|| || g}D]}||vs|j5||dd}|rd$jA|nd}t|||%S)&Nrerqrsrr!laughterT)shortwords)rgN@g333333?r(r)rrpitchrrvoicedr,r.g?rg?rr g ףp= ?r$g?g333333?rrr"c3&K|]}|v ywr0rrRrSr s rrTz&ProsodyReader._read..(sF3EaT 3E)rr rrr#c3&K|]}|v ywr0rrs rrTz&ProsodyReader._read..0s!*)&'qDy)r)r r$rrr"rr#z, )r r r)!rlrtsplitr)rvrzr{astyperx _MIN_SPEECH_Sr rYr~rstdr _MIN_VOICED percentilemedianroundr-r_SEED_MIN_LOUD_SEED_NrrrrGrrranyjoin)r3r|rrkrr!n_wordsrfrmsrrp25p75r pitch_medrseedabler loud_hitrr"deviatedseenrSr s @rrzProsodyReader._reads4=tzr((*+ <bjj)AQ)F!G$!NOPD m #w!|zJ<*d9KLL4'!BC CWns8S11% ;?99uTXXZ499;#567#U# :: $}}URH5HC39%Fbii./I!% %FY:uT1~)2U9%'-'9eFmt$Nc%**oDGG  >)GejjK.G 77W  6%%c64>!):tD4*5I Ir5I I  KK #HD4??S3H,HI a<T^^d** Hf=t~~,, F#  $"3"38S!2!2S!888$//C"77 T..445t44  J' F# CF3EFF KK  __$(>!> KK *)**H  ! !#vtT :A} ABQx"&tyyDDt599r)rFNoner0) r| np.ndarrayrr)rkstrrz float | NonerFr )rFr )rrrrrrrrrrrrrs2M= #'+2"W:rr)rFr')r_z 'Prosody'rFr)rkrrFrG)r|rrFr).r __future__rrrf threading dataclassesrrnumpyrvrrurrrrrrrrrcompilerir r1r6r:r=r@rCrEr]r^rOrWr'Lockr\rJrKrcrlr~rrrrrrs2<# (         BJJM N  .. .$"*$ I$ J    !3#s!4$NsD9  $@@@4Y^^  k  $% 2-