o 5i L@sldZddlZddlZddlZddlmZddlm Z m Z m Z ddl m Z eGdddZGdd d ZdS) a Audio Polisher - Segment boundary cleanup for transcription quality. Two-phase polishing: Phase 1 (original): Remove CUT SPEECH ARTIFACTS from VAD boundaries - Detects burst -> gap -> speech pattern and removes the burst - Silence at boundaries preserved as natural padding Phase 2 (new): TIGHTEN boundaries to actual speech onset/offset - Energy-based detection of where speech actually starts/ends - Cuts to those points with configurable silence margin (~50ms) - Ensures clean, crisp boundaries for TTS training data - Prevents segments starting/ending mid-voice-energy Combined result: artifact-free + precisely-bounded audio segments. No ML models - pure signal processing, runs in <10ms/segment. N) dataclass)OptionalTupleList)Pathc@seZdZUdZeed<eed<eed<dZeed<dZ eed<dZ eed <dZ eed <dZ eed <dZ eed <dZeed <dZeed<dZeed<dZeed<dZeed<dZeed<eddZddZdS) PolishResultz/Result of audio polishing for a single segment. input_path output_path was_modifiedstart_trimmed_msend_trimmed_msclean start_quality end_qualityoriginal_duration_mspolished_duration_mssnr_dbrms_dbpeak_dbF is_clippedvolume_adjustedvolume_gain_dbcCs |j|jSN)r r )selfr;/home/ubuntu/maya3_transcribe/src/backend/audio_polisher.pytotal_trimmed_ms,s zPolishResult.total_trimmed_msc Cst|jj}|js d|d|jdd|jdd|jd|j Sg}|jdkr6| d |jdd |jd |j dkrJ| d |j dd |jd |j rX| d |j ddd |}d|d|jdd|jdd|d|jdd S)Nz z : SKIP | z.0fz ms | SNR:z.1fz dB | start:z end:rzstart -zms ()zend -zgain +dBz | z : POLISHED | zms -> zms | z | SNR:)rrnamer rrrrr appendr rrjoinr)rr partschangesrrrsummary0s8    zPolishResult.summaryN)__name__ __module__ __qualname____doc__str__annotations__boolr floatr rrrrrrrrrrpropertyrr%rrrrrs(               rc@seZdZdZ                  d+ddZddZddZddZddZddZ dd Z d!d"Z d,d$d%Z d-d'd(Z d-d)d*Zd&S). AudioPolisherz Two-phase audio segment polisher for transcription quality. Phase 1: Conservative artifact removal (burst-gap-speech pattern detection) Phase 2: Energy-based boundary tightening (speech onset/offset + silence margin) $@@i@b@.@Y@ffffff?333333?@>@r@2(@Gz?TI@flaccCs||_||_||_||_||_||_||_||_| |_| |_ | |_ | |_ | |_ ||_ ||_||_||_||_||_||_||_||_||_dSr)frame_mshop_msmax_start_trim_msartifact_search_ms min_gap_msmax_end_trim_ms min_burst_ms max_burst_msenergy_gap_factorzcr_artifact_thresholdmin_dynamic_range_dbmin_silence_pad_msmin_remaining_ms min_trim_ms fade_in_ms fade_out_ms target_rms_dbsnr_boost_threshold_db max_gain_dbclipping_thresholdtighten_boundaries_enabledsilence_margin_ms output_format)rr@rArBrCrDrErFrGrHrIrJrKrLrMrNrOrPrQrRrStighten_boundariesrUrVrrr__init__Ms. zAudioPolisher.__init__c Cst|j|d}t|j|d}tdt|||d}t|}t|}t|}t|D]P} | |} t| |t|} || | } t t | d} dt | d|| <t| dkrxt t tt| dt| || <| |d||| <q2|||fS)z5Compute per-frame energy (dB) and zero-crossing rate.绽|=)intr@rAmaxlennpzerosrangeminsqrtmeanlog10sumabsdiffsign)raudiosr frame_samples hop_samplesn_framesste_dbzcr frame_timesistartendframermsrrr_compute_framess$        zAudioPolisher._compute_framescCsf|jd}t||kd}t|dkrdSt|d}t|d}||}||jkr-dS||j|} tdt|} ||d| } t| | krLdS||| k} g} d}d}t t| D]0}| |rk|skd }|}q^| |s|rd}||||||d}||j kr| |||fq^|r||d |||d}||j kr| |t| |f| sdS| d}|ddkrdS|d t|krd S|d|d}t|dkrt ||nd}||d}||}|d|j ks|d|jkrd|d|j krd fSdfS|d}||jkr||jkrd S||jkr#|dfS|d|j dkr1|dfSd S)ar Detect CUT SPEECH artifact at segment start. ONLY acts when: burst (cut speech) -> gap (silence) -> real speech. Trims at gap START = removes burst, keeps gap as silence padding. Does NOT trim if segment starts with silence (that's good padding). Does NOT trim if segment starts with continuous speech (natural start). rYr)NrUNFTrZ)N marginal_keptr~rr[artifact_trimmedg?)rCrawherer` percentilerJrHrdallrcrDr!rfrBrMrKrI)rrqrrrs search_limitsinfsldr gap_threshn_checkfirst_frames_energybelowgapsin_gapgsrtdurgpremzcrgap_start_framettgap_msrrr_detect_leading_artifactsl               z&AudioPolisher._detect_leading_artifactcCsZt|d}t|d}||}||jkrdS||j|}||k}t|} td| } t|| dr8dSt|d} t| dkrGdS| d} td| } | dkr|| | | }t|}|d kr| }t | d ddD] }||rx|}qo|| ||d }|j |kr|j krndS||}||d }||j kr||j kr|d d fSdS)z Detect CUT SPEECH artifact at segment end. ONLY acts when: sustained speech -> silence gap -> isolated burst at end. r{r|)NrFrzNrr} g333333?rZrYrF)NrT)rarrJrHr`rdrrrfrcrFrGrErM)rrqrs total_durrrrthreshabovenrai last_abovelookback preceding sustained burst_startbi burst_durr actual_trimrrr_detect_trailing_artifactsB         z'AudioPolisher._detect_trailing_artifactc Cs\|d}d}d}t|D]\}}||kr)|d7}||kr(||d} || Sq d}q dS)as Find precise speech onset: first sustained energy rise above noise floor. Uses hysteresis: speech starts when energy exceeds threshold for 3+ consecutive frames (~15ms at 5ms hop). Prevents transient noise spikes from being misidentified as speech onset. Returns: onset time in seconds, or None if no clear onset found. r8rzrrZN) enumerate) rrqrsnoise_floor_dbr consec_neededconsecrte onset_framerrr_find_speech_onsets   z AudioPolisher._find_speech_onsetc Csv|d}d}d}tt|dddD]&}|||kr6|d7}||kr5t||dt|d}||Sqd}qdS)z Find precise speech offset: last sustained energy above noise floor. Mirrors onset detection, scanning backward. Returns: offset time in seconds, or None if no clear offset found. r8rzrrZr}N)rcr`rd) rrqrsrrrrrt offset_framerrr_find_speech_offset-s  z!AudioPolisher._find_speech_offsetcCst||}t|d}||||}||||}|dus"|dur$dS|jd} td|| } t||| } | d} || d} | dkrK| dkrKdS| | d}||jkrXdSt | |}t | |}|||| | fS)a Tighten segment boundaries to actual speech onset/offset. After artifact removal, precisely locates where speech energy begins and ends, then cuts to those points with silence_margin_ms padding. Ensures segments don't start/end mid-voice-energy. Returns: (tightened_audio, start_trim_ms, end_trim_ms) or (None, 0, 0) if no tightening needed rN)Nr r g@@r rYr9) r`rarrrrUr_rdrLr^)rrlrmrqrsr noise_flooronsetoffset margin_sec new_startnew_end start_trimend_trim new_dur_msssesrrr_tighten_to_speechBs&       z AudioPolisher._tighten_to_speechcCst|dt|dS)z&Estimate SNR from energy distribution.Zr)rar)rrqrrr _measure_snriszAudioPolisher._measure_snrFcCs|}|r-|jdkr-t|j|d}|dkr-|t|kr-|d|tdd|9<|rW|jdkrWt|j|d}|dkrW|t|krW|| dtdd|9<|S)z9Apply smooth fade-in/fade-out to prevent click artifacts.rrYNrZ)copyrNr^r`ralinspacerO)rrlrmfade_infade_outoutrrrr _apply_fadems zAudioPolisher._apply_fadeNc)Cst|\}}t|jdkr|jdd}t||d}tt|d}tt|}dt |d} dt |d} ||j k} ||j krVt ||d||| | | dS| ||\} } }|| }|| | |\}}|| |t||\}}}|d ur|dnd }|d urt|||dnd }||||j krd \}}d \}}|d krdnd }|d krdnd }d}d}d }||jkr| |jkrt|j| |j}|dkrd}nd }|}|d urt||nd}|d urt||nt|}|||}|d up|d u}|jrPt||d|j krP| ||\}}}|||||\} }!}"| d urP| }||!7}||"7}|d krD|!dkrDd}|d krP|"dkrPd}|j|||dk|d}|r{d|d}#||#}tt|}$|$dkr{|d|$}|p|dkp|dkp|}%|%st ||d|||||| | | d St||d}&|d ur|rtj|ddtj|t|j}nt|j }'t|j!pd|j"}(t#t|j$|'d|(}t%|||t ||d||||||&|| | | ||dS)a% Polish a single audio segment (two-phase). Phase 1: Remove cut speech artifacts (burst-gap-speech pattern) Phase 2: Tighten boundaries to actual speech onset/offset + silence margin Already-clean segments pass through unchanged (was_modified=False). rZ)axisrYr[r\r]F)rr r rrrrrNr NN)r r rr~g?Trboundary_tightened)rrrr=) rr r rrrrrrrr)exist_ok. _polished)rr r r r rrrrrrrrrr)&sfreadr`shaperfrarer_rirgrSrLrryrrrrQrPrdrRrr^rTrrosmakedirspathr"rr stemsuffixrVr*parentwrite))r audio_pathr output_dirrlrmorig_msrms_valpeakrrrrqrrftrs_secs_qe_sece_qneeds_fade_outs_mse_msvol_adjvol_gainpolrrphase1_modifiedste2_ft2 tightenedt_startt_endrmx needs_modpol_msrextrrrpolishzs               zAudioPolisher.polishc Cs|dur tj|d}hd}g}|D]}|t|d|qt|ddd}|r3|d|}tdt|d g}d }|D]} |j t | |d } | | | j rZ|d 7}t| qCtd |dt|dt||d|S)z)Polish all audio segments in a directory.Npolished>.m4a.mp3.ogg.wav.flac*cSs|jSr)r )xrrrsz0AudioPolisher.polish_directory..)keyz[Polisher] Processing z segments...r)rrZz[Polisher] Done: /z modified, z already clean)rrr"extendrglobsortedprintr`rr*r!r r%) r input_dirr max_filesextsfilesrresultsmodifiedfrrrrpolish_directorys,    zAudioPolisher.polish_directory)r0r1r2r3r4r3r0r5r6r7r8r9r:r0r1r0r;r<r<r=Tr>r?)FFr)r&r'r(r)rXryrrrrrrrrrrrrrr/EsH  3M4'  ~r/)r)rnumpyra soundfiler dataclassesrtypingrrrpathlibrrr/rrrrs  +