j`j ddlZddlZddlZddlmZddlmZddlZddlm Z ddl m Z ddl mZmZdd d d d d d dd dZdZeejdz dz ZdZdZdZdZgdZdddegZgdZgdZdddd ddd!ddd"d#dd$d%dd&d'dd(d'dd)d*dd+d,dd- Zd.d/dd0dd1d2Zgd3d4d5d6d3d7d8d9Z d:Z!d;Z"dd?d@e"e"e$dAdBgZ%dCe&dDe&fdEZ'dFe&dDe&fdGZ(dHe&dDee&fdIZ)dDee&fdJZ*dHe&dDee+fdKZ,dHe&dDe-e&fdLZ.dHe&dDe-e&fdMZ/dHe&dDe-e0e&e&ffdNZ1dHe&dDe-e0e&e&ffdOZ2dHe&dDe+fdPZ3dDe4fdQZ5dHe&dDe-e+fdRZ6GdSdTZ7dS)UN)Path)Optional) mps_device_or) files_locator)TTS_MONOLOGUE_PROMPTTTS_QWEN3_DIALOGUE_PROMPTz$Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoicezqwen3_tts_customvoice.json)repo config_filez$Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesignzqwen3_tts_voicedesign.jsonzQwen/Qwen3-TTS-12Hz-1.7B-Basezqwen3_tts_base.json)qwen3_tts_customvoiceqwen3_tts_voicedesignqwen3_tts_basez qwen3_tts_generation_config.jsonqwen3configsqwen3_tts_text_tokenizerqwen3_tts_tokenizer_12hzz$qwen3_tts_tokenizer_12hz.safetensorszDeepBeepMeep/TTS)z merges.txtz vocab.jsonztokenizer_config.jsonpreprocessor_config.jsonz config.jsonzconfiguration.jsonr) autochineseenglishjapanesekoreangermanfrenchrussian portuguesespanishitalian) serenavivianuncle_furyanaidenono_annasoheeericdylanz(Bright, slightly edgy young female voiceChinese)stylelanguagezWarm, gentle young female voicez-Seasoned male voice with a low, mellow timbrez8Youthful Beijing male voice with a clear, natural timbrezChinese (Beijing Dialect)z:Lively Chengdu male voice with a slightly husky brightnesszChinese (Sichuan Dialect)z-Dynamic male voice with strong rhythmic driveEnglishz/Sunny American male voice with a clear midrangez9Playful Japanese female voice with a light, nimble timbreJapanesez*Warm Korean female voice with rich emotionKorean) r rr!r'r&r"r#r$r%zMax duration (seconds)z Max DurationiX)labelnameminmax incrementdefaultAABABDzVoice cloning of 1 speakerz5Voice cloning of 2 speakers (Speaker 1 and Speaker 2)z0Voice cloning of 3 speakers (Speaker 1, 2 and 3)r8r6) selectionlabelsletters_filterr4auto_split_every_sgg@gV@zAuto Split s (0 = Auto, 5-90)zAuto Split Every sfloatg?)idr/r0typer4r1r2incvaluereturncR|ddS)N_ )replacetitle)rAs @C:\Users\johnn\OneDrive\Documents\AI\models\TTS\qwen3_handler.py_format_qwen3_labelrIs" ==c " " ( ( * **r0cvt|}t|}|s|Sg}|dd}|dd}|r|||r|||s|S|dd|dS)Nr)r*z (z; ))rIQWEN3_TTS_SPEAKER_METAgetlowerappendjoin)r0r/metapartsr)r*s rH_format_qwen3_speaker_labelrUs  % %E ! % %djjll 3 3D  E HHWb ! !Exx B''H  U X   * *tyy'' * * **rJbase_model_typect|}|dSt|dz }|rt |ndS)Nr )QWEN3_TTS_VARIANTSrO_QWEN3_CONFIG_DIRis_filestr)rVvariant config_paths rHget_qwen3_config_pathr^sR $$_55Gt#gm&<3{   $>rJcjttz }|rt|ndSN)rYQWEN3_TTS_GENERATION_CONFIGrZr[)r]s rH get_qwen3_generation_config_pathrbs1#&AAK*2244 >3{   $>rJct|}|sdSt|dd5}tj|cdddS#1swxYwYdS)Nrzutf-8)encoding)r^openjsonload)rVr]handles rHload_qwen3_configrjs'88K t k3 1 1 1!Vy  !!!!!!!!!!!!!!!!!!sAA A ct|}|ttS|didi}d|D}dgt d|Dz}|S)N talker_configcodec_language_idc>g|]}d|v|S)dialectrP.0r0s rH z'get_qwen3_languages..s*SSS$Ydjjll5R5R5R5R5RrJrc6h|]}|Srprqs rH z&get_qwen3_languages..s "F"F"FD4::<<"F"F"FrJ)rjlistQWEN3_TTS_LANG_FALLBACKrOkeyssorted)rVconfiglang_map languagess rHget_qwen3_languagesr~s  / /F ~+,,,zz/2..223FKKHSS(--//SSSI6"F"FI"F"F"FGGGI rJc<t|}|ttSt|didi}t d|D}|pttS)Nrlspk_idc6h|]}|Srurprqs rHrvz%get_qwen3_speakers..s 999tzz||999rJ)rjrwQWEN3_TTS_SPEAKER_FALLBACKrOryrz)rVr{speakerss rHget_qwen3_speakersrs  / /F ~.///FJJ3377"EEJJLLMMH99999::H  7t6777rJc4dt|DS)Nc0g|]}t||fSru)rI)rrlangs rHrsz.get_qwen3_language_choices..s& _ _ _$  & & - _ _ _rJ)r~rVs rHget_qwen3_language_choicesrs _ _:Mo:^:^ _ _ __rJc4dt|DS)Nc0g|]}t||fSru)rUrqs rHrsz-get_qwen3_speaker_choices..s& f f f$ ( . . 5 f f frJ)rrs rHget_qwen3_speaker_choicesrs f fBTUdBeBe f f ffrJc iddddddddddd dd dd d d dd|gdttdddtddddddddd|dkrddgndgdddddd}|d kr=t|}|r|dnd }i|t ||d!d"d#d$d%d&d'd(d)d*d+S|d,kr"i|d-d.d/d0t |d1d2d"d3d4d)d*d5S|dkrti|id6t |d1d2d"d7d8d9d:d*d;dddd?tt d@dAtDdBtdCddDdEdFdGdHdIdJdKdLdMdNdOdPdQS|S)RN audio_onlyT image_outputsFsliding_windowguidance_max_phasesrno_negative_promptinference_steps temperatureimage_prompt_types_allowedrLsupports_early_stop profiles_dirduration_slider top_k_slider!text_prompt_enhancer_instructionstext_prompt_enhancer_max_tokensiprompt_enhancer_button_labelWritecompileparent_model_typerTT1z A Speech based on current Promptz5A Dialogue between two People based on current Prompt)rr)r9r:r4) lm_decoderprompt_enhancer_defr Speaker)choicesr4r/aXSpeak the Text Prompt (`prompt`) with one of Qwen3's built-in voices, selected in **Speaker** (`model_mode`); each label gives the voice's style and native language. The optional **Instruction** (`alt_prompt`) adjusts the delivery, for example `calm, friendly, slightly husky`. Max duration caps the output; give the text enough time to finish.zWrite only the words to speak, with natural punctuation. Put delivery directions in Instruction rather than in the text, for example Instruction `excited, fast-paced` with the text "We did it! The launch window is open."zBuilt-in voices: speaker = `model_mode` (labels give style and language). `prompt` = words to speak; `alt_prompt` = optional delivery instruction. Max duration caps output.zyWrite only the words to speak, naturally punctuated. Put delivery in `alt_prompt`, e.g. 'calm, friendly, slightly husky'.zInstruction (optional)zcalm, friendly, slightly husky)r/ placeholderlines) model_modesinfos prompt_infos deepy_infosdeepy_prompt_infos alt_promptr aOGenerate speech by combining the exact words in Text Prompt (`prompt`) with a voice description in Voice instruction (`alt_prompt`). The description defines the voice and delivery for that speech. Select the speech language with Language (`model_mode`), or use Auto. Max duration caps the output; give the script enough time to finish.zVoiceDesign combines spoken text (`prompt`) and a voice/delivery description (`alt_prompt`). Language is `model_mode` (or Auto); Max duration caps speech length.zPut only the words to speak in `prompt`, naturally punctuated in the target language. Describe age, timbre, accent, pace and emotion in `alt_prompt`, e.g. 'A low, warm male voice, speaking slowly with dry amusement.' Keep one consistent voice description.aPut only the words to be spoken in `prompt`, with natural punctuation and spelling in the target language. Put age, timbre, accent, pace and emotion in `alt_prompt`, for example: "An older man with a low, slightly rough voice, speaking slowly with dry amusement." Example spoken text: "Welcome aboard. I promise the robot is doing most of the driving." Keep one consistent voice description for the utterance.rLanguagezVoice instruction+young female, warm tone, clear articulation)rrrrrrrrz?Reference transcript(s) (optional, multi-speaker: one per line)z\Speaker 1 reference transcript Speaker 2 reference transcript Speaker 3 reference transcriptpause_between_sentencespreserve_empty_prompt_linesany_audio_promptaudio_prompt_choicesaudio_prompt_type_sourcescustom_settingsc6g|]}|Sru)copy)rrones rHrsz'get_qwen3_model_def../s PPPs PPPrJ"text_prompt_enhancer_instructions1 text_prompt_enhancer_max_tokens1audio_guide_labelzSpeaker 1 reference voiceaudio_guide2_labelzSpeaker 2 reference voiceaudio_guide3_labelzSpeaker 3 reference voiceraCClone one to three voices from short reference recordings and speak the Text Prompt (`prompt`) with them. - **Voice cloning of 1 speaker:** the whole text is spoken with the Speaker 1 reference voice. - **Voice cloning of 2 or 3 speakers:** write the script as `Speaker 1:`, `Speaker 2:` (and `Speaker 3:`) lines. Each speaker's lines use its own reference voice, in upload order, and the script must use exactly as many speaker numbers as reference voices. **Reference transcript(s)** (`alt_prompt`, one line per speaker in order) give the exact words spoken in each reference. With a transcript Qwen3 imitates the voice more closely; without one it only uses the overall character of the voice. Use short, clean references of a single speaker. **Language** (`model_mode`) selects the speech language, or Auto. **Pause between sentences** adds silence between turns and split segments, **Auto Split** cuts long text into segments of about that many seconds (0 keeps the automatic segmentation), and Max duration caps the assembled audio. Early Stop returns the segments already completed.raWrite only the words to speak, with natural punctuation in the target language; the delivery follows the reference voice. For a dialogue, start every turn with its speaker number: ```text Speaker 1: We should leave before the rain gets heavier. Speaker 2: Give me one minute, I still need my jacket. Speaker 3: Hurry, the bus is already at the corner. ``` Use as many speaker numbers as reference voices, and keep stage directions out of the text.raw`prompt` = words to speak. `audio_prompt_type`: `A` clones `audio_guide`; `AB` / `ABD` make a dialogue with `audio_guide2` (and `audio_guide3`), and the script must use exactly 2 / 3 `Speaker N:` IDs in sample order. `alt_prompt` = optional exact transcripts of the samples, one line per speaker, for closer cloning. Language = `model_mode` or Auto; Max duration caps output.rzWrite only the words to speak, naturally punctuated; delivery follows the reference voice. Dialogue: one `Speaker N:` line per turn, e.g. 'Speaker 1: Ready? Speaker 2: Almost.' No stage directions.) dictQWEN3_TTS_DURATION_SLIDERrrrr#QWEN3_TTS_AUDIO_PROMPT_TYPE_SOURCESQWEN3_TTS_CUSTOM_SETTINGSr )rVcommonrdefault_speakers rHget_qwen3_model_defrsd % q  d  5  t %b t ) 4 9::  ,-A *3 ' 5!" -#$,;?O,O,Oc4[[VYUZ;Q $ $ 'F8111%o66)19(1++r  5_EE*" p{J#^1?    "111  g#dx5oFF!# -L    "***   5oFF!#  Z  &t  *4    #D (.Q)R)R! " PP6OPPP# $ 12K% & /' ( !<) * !"=+ , !"=- . `/ 0 h1 2 U3 4 !#k5   8 MrJcRttddS)Nr4r.)intrrOrurJrHget_qwen3_duration_defaultr=s! (,,Y;; < <78  %!? @9:   rJc.eZdZedZedZedZedZedZedZ eddZ e dd Z ed Z ed Z edZedZdS)family_handlerc*ttSr`)rwrXrurJrHquery_supported_typesz$family_handler.query_supported_typesQs&'''rJc iifSr`rururJrHquery_family_mapsz family_handler.query_family_mapsUs 2v rJcdS)NttsrururJrHquery_model_familyz!family_handler.query_model_familyYsurJc ddiS)Nr)iTTSrururJrHquery_family_infosz!family_handler.query_family_infos]s }%%rJcdS)N qwen3_ttsrurs rH get_lora_dirzfamily_handler.get_lora_diras{rJcdddlm}|dkrdn |dkrdnd}it|d||d S) Nr) specialitiesr)z voice cloningzspeech dialoguer )z voice designruT)device_explicitr)models.help_sharedrr)rV model_defr strengthss rHquery_model_defzfamily_handler.query_model_defes|333333""Q&&$21$5! .   1>*;5QQQL#0 ##++ e!4!455#*    x~. 8/ 6 6 B'/'@'FD# $  ql q 0 0 0 0 0 0/@@K"#%7<<D]D]D]#^#^  n-=USSSgWg \5KbTYTbdo p p p$I[6f}ooafggggrJcNd|vrd|d<|dkrt|}d|r|dndd}n|dkrddd}n|dkrd dd d }nddd}|D]\}}||||dkr}t|d d pd }dd |Ddvr%d dd|Dz|d <|dkr|ddr d|d<dSdSdS)NrrLr r)audio_prompt_type model_moder rrr6?)rr pause_secondsrc3"K|] }|dv|V dSr8Nrurrflags rH z.family_handler.fix_settings..s&KKTU]]t]]]]KKrJr5c3"K|] }|dv|V dSr!rur"s rHr$z.family_handler.fix_settings..s.@y@y$gksxgxgxgxgxgxgx@y@yrJgQ@rF2top_k)ritems setdefaultr[rOupperrR) rVsettings_versionr ui_defaultsrdefaultskeyrArs rH fix_settingszfamily_handler.fix_settingss { * *(*K % 5 5 5)/::H%'-5=hqkk2HH 7 7 7%'$HH 0 0 0%($!$HH&($H#..** / /JC  " "3 . . . . . . . #KOO4G$M$M$TQT U U [ [ ] ] wwKK(9KKKKKSeee36@y@yRc@y@y@y9y9y3y /0 d " "}}^U33 *') G$$$ # " * *rJc||dkrJt|}|r|dnd}|d|dtdddddddd dS|d kr/|dd d tdddddddd dS|d kr0|d d dtddddddddd dSdS)Nr rrLrg?r&FG) rrrduration_secondsrepeat_generation video_lengthnum_inference_stepsnegative_promptrr'multi_prompts_gen_typer rrrr6r) rrrr2rr3r4r5r6rr'r7)rupdater)rVrr,rrs rHupdate_default_settingsz&family_handler.update_default_settingssC 5 5 5)/::H-5=hqkk2O   )+"1"$(B(D(D)*$%+,')#&.2      F 5 5 5   )+"("O(B(D(D)*$%+,')#&.2      F . . .   ),"("$(B(D(D%()*$%+,')#&.2        / .rJc4|dkr}|2tt|dkrdS|dd}|sdSt |}||vrd|dSdS|d kr8|2tt|dkrd SdS|d krO|2tt|dkrd St|d dpd}|ddSt|}tjd|tj du}d|vr|ddSd|vrdnd} ttj d|tj } | s| d| dStd| D} t| | kr | d| d| dSn|rd SdSdS)!Nr rz2Prompt text cannot be empty for Qwen3 CustomVoice.rrLz.Please select a speaker for Qwen3 CustomVoice.zUnsupported speaker 'z'.r z2Prompt text cannot be empty for Qwen3 VoiceDesign.rz7Prompt text cannot be empty for Qwen3 Base voice clone.rr6 audio_guidez.Qwen3 Base requires Speaker 1 reference audio.zSpeaker\s*\d+\s*:)flagsB audio_guide2z4Two-speaker mode requires Speaker 2 reference audio.DrrzSpeaker\s*(\d+)\s*:z@-speaker mode requires prompt lines using Speaker 1: to Speaker z: cRh|]$}t|d%S)r)rgroup)rrms rHrvz.&s(%O%O%O!c!''!**oo%O%O%OrJz-speaker mode requires exactly z( speaker IDs. Use Speaker 1: to Speaker z?:. For headless settings, keep 'multi_prompts_gen_type' = 'FG'.zASpeaker-tag dialogue requires a multi-speaker voice cloning mode.) rr[striprOrrPr*research IGNORECASErwfinditerrz) rVrinputs one_promptspeakerrr prompt_texthas_speaker_syntax speaker_countspeaker_matches speaker_idss rHvalidate_generative_promptz)family_handler.validate_generative_prompts 5 5 5!SZ)>)>)@)@%A%AQ%F%FKKjjr22G HGG)/::H}}h..:w::::4 5 5 5!SZ)>)>)@)@%A%AQ%F%FKK4 . . .!SZ)>)>)@)@%A%AQ%F%FPP #FJJ/BC$H$H$OC P P V V X X zz-((0GGj//K!#+?TVTa!b!b!bjn!n '''::n--5QQ%(,=%=%=1 "&r{3I;^`^k'l'l'l"m"m&+~~mz~~~~$%O%O%O%O%OPP {##}44(WWWWANWWW5 $ [ZZ4trJc|dkrdS|dd}|dSt|tsdS|td}|dSt|tr^|}t |dkr7|tdt |dkr|nd|d<dS t|trtt|}n=#t$r0dttdttdcYSwxYw|tkr7|tdt |dkr|nd|d<dS|tks |tkr-dttdttdS||t<||d<dS) Nrrz%Custom settings must be a dictionary.rz,Auto Split Every s must be a number between z and z seconds.z#Auto Split Every s must be between )rOrrQWEN3_TTS_AUTO_SPLIT_SETTING_IDr[rCrpopbool ValueErrorr= Exceptionr QWEN3_TTS_AUTO_SPLIT_MIN_SECONDS QWEN3_TTS_AUTO_SPLIT_MAX_SECONDS%QWEN3_TTS_AUTO_SPLIT_DISABLED_SECONDS)rVrrHr raw_valueauto_split_secondss rHvalidate_generative_settingsz+family_handler.validate_generative_settings2s8 . . .4 **%6==  "4/400 ;::#''(GNN  4 i % % !))I9~~""##$CTJJJ?B??S?SVW?W?WOO]a()t )T** # ll"!&y!1!1     p788pp?BCc?d?dppp     !F F F    ? F F F;>;O;ORS;S;SY]F$ %4 !A A A!$DDDp788pp?BCc?d?dppp  rjs  $$$$$$,,,,,,MMMMMMMM 73 73 0,  ADNN**,,3g= I9!;%K"#"""& $       < 3 A L/ N/   A   C M >C%%N&   $## )EA  ' '##7(+%#& #' .0$84/   +s+s+++++c+c++++"?3?8C=?????(3-???? !s!x~!!!!c88S 8888``U38_8M````gsgtE#s(O7Lgggg``````F=C==== C DJ    OOOOOOOOOOrJ