.j[6T%SSKrSSKrSSKrSSKJr SSKJrJrJ r J r J r J r SSK Jr SSKrSSKrSSKr\ \\ \R*\44r\ \\ \4rSrSrSrSr/S Qr\ \\S 'S rS r S \S\4Sjr!S \SS4Sjr"S\S\ \4Sjr#S\S\$4Sjr%S\S\$4Sjr&S\S\'4Sjr(S\S\ \R*\44Sjr)S\R*S\R*4Sjr*S\R*S\R*4Sjr+S\S\R*4Sjr,S\ \\ \4S\ \R*4Sjr-S \ \S!\S\\ \4S"jr.\"S#S$9"S%S&55r/S3S'\R*S(\S)\0S*\0S+\0S\ \ \R*\044 S,jjr1S4S-jr2S5S.\S/\ \S\ \\44S0jjr3S1\ \S\4S2jr4g)6N) dataclass)AnyIterableListOptionalTupleUnion)urlparsei>ig?)ChineseEnglish CantoneseArabicGermanFrenchSpanish Portuguese IndonesianItalianKoreanRussianThai VietnameseJapaneseTurkishHindiMalayDutchSwedishDanishFinnishPolishCzechFilipinoPersianGreekRomanian Hungarian MacedonianSUPPORTED_LANGUAGESz z language languagereturncUc [S5e[U5R5nU(d [S5eUSSR5USSR 5-$)a Normalize language name to the canonical format used by Qwen3-ASR: first letter uppercase, the rest lowercase (e.g., 'cHINese' -> 'Chinese'). Args: language (str): Input language name. Returns: str: Normalized language name. Raises: ValueError: If language is empty. Nzlanguage is Nonezlanguage is empty) ValueErrorstrstripupperlower)r+ss 8/mnt/workspace/git/Qwen3-ASR/qwen_asr/inference/utils.pynormalize_language_namer6IsZ+,, H A ,-- Ra5;;=1QR5;;= ((cBU[;a[SUS[35eg)z Validate the language is supported. Args: language (str): Canonical language name. Raises: ValueError: If unsupported. zUnsupported language: z . Supported: N)r*r/)r+s r5validate_languager9_s-**1(=I\H]^__+r7xc6[U[5(aU$U/$N) isinstancelist)r:s r5 ensure_listr?msq$''q0aS0r7r4c[U5nURS;=(a [UR5$![a gf=f)N)httphttpsF)r schemeboolnetloc Exception)r4us r5is_urlrHqs? QKxx,,?ahh? s58 AAcjURS5(agSU;aSU;a[U5S:agg)Nz data:audioT/\F) startswithlen)r4s r5is_probably_base64rOys1||L!! 1 QCFSL r7b64cSU;a9UR5RS5(aURSS5Sn[R"U5$)N,zdata:r.)r1rMsplitbase64 b64decode)rPs r5decode_base64_bytesrVsE czciik,,W55iiQ"   C  r7c[U5(ap[RRU5nUR 5nSSS5 [ R "W5n[R"USSS9upESSS5 Ok[U5(aC[U5n[ R "U5n[R"USSS9upESSS5 O[R"USSS9upE[R"W[RS9n[W5nXE4$!,(df  N=f!,(df  NP=f!,(df  Na=f)Nfloat32F)dtype always_2d)srmonorY)rHurllibrequesturlopenreadioBytesIOsfrOrVlibrosaloadnpasarrayrXint)r:resp audio_bytesfaudior[s r5load_audio_anyrns ayy ^^ # #A &$))+K' ZZ $eDIE% $ A  )!, ZZ $eDIE% $LLt%8  JJuBJJ /E RB 9' & $ $% $s#DD.2D? D+. D<? E rmcXURS:XaU$URS:XaqURSS::a,URSURS:a URn[R"USS9R [R 5$[SUR35e)Nr.raxiszUnsupported audio ndim=)ndimshapeTrgmeanastyperXr/)rms r5to_monorzs zzQ  zzQ ;;q>Q 5;;q>EKKN#BGGEwwu2&--bjj99 .uzzl; < Y D MM4$%'",-G-/F3 771: C O+Cq1c(SIPPQSQ[Q[\A q#h  F Mr7cB^SnSU4SjjmUnU"X15nT"X5nU$)Nc,/nSn[U5nX4:apSnX5-U:a)XU-X:XaUS- nX5-U:aXU-X:XaMXQ:aURX5 X5- nOURXX5-5 X5- nX4:aMpSRU5$)Nrr.)rNrjoin)r4threshresrncounts r5fix_char_repeats4detect_and_fix_repetitions..fix_char_repeatsPs  FeE)a-A%iLAD$8 )a-A%iLAD$8~ 14   1qw<( ewws|r7c~>[U5nUS-nX4:aU$Sn/nXSU- ::aSn[SUS-5HnXXU--U:a OXXX-n Sn [SU5Hn X[U--n X X-U :wdMSn O U (dMIUn XQU--nX-U::a'XX-U :XaU S- n X- nX-U::a XX-U :XaMURU 5 URT"XSX55 UnSn O U(aO"URX5 US- nXSU- ::aMW(dURXS5 SRU5$)NrprFr.Tr)rNrrr)r4rrrmin_repeat_charsrresultfoundkpatternrrep start_idx total_rep end_indexfix_pattern_repeatss r5r7detect_and_fix_repetitions..fix_pattern_repeatsas F!A: H '''E1gk*6z>A%ac( F+C !!G I9;/7: % , 5 &I !QJI#-1,Y[1IW1T!Q ! $-1,Y[1IW1TMM'*MM"5a mV"UVA E-+0 ad#Q='''@ MM!B% wwvr7r)text thresholdrtext_rawrs @r5detect_and_fix_repetitionsrOs."*XH H 0D t /D Kr7raw user_languagecUcg[U5R5nU(dg[U5nU(aX4$UnSn[U;nU(aUR [S5up4OSUR54$UR 5nSU;aUR5nU(dgSU4$SnUR 5Hzn U R5n U (dMU R 5n U R[5(dMHU [[5SR5n U (a [U 5n O XR54$)aP Parse Qwen3-ASR raw output into (language, text). Cases: - With tag: "language Chinese...." - With newlines: "language Chinese\n...\n...." - No tag: treat whole string as text. - "language None": treat as empty audio -> ("", "") If user_language is provided, language is forced to user_language and raw is treated as text-only (the model is expected to output plain transcription without metadata). Args: raw: Raw decoded string. user_language: Canonical language name if user forced language. Returns: Tuple[str, str]: (language, text) N)rrrr.z language none) r0r1r _ASR_TEXT_TAGrSr3 splitlinesrM _LANG_PREFIXrNr6) rrr4 meta_part text_parthas_tag meta_lowertlanglinelowvals r5parse_asr_outputrs). { CA "1%AIIq G ww}a8 91779}"J*$ OO 1u  D$$&zz| jjl >>, ' 's<()*002C.s3 ' " ""r7langsc/nSnUH?nU=(d SR5nU(dM%X2:XaM,URU5 UnMA SRU5$)a) Merge per-chunk languages into a compact comma-separated string, keeping order and removing consecutive duplicates and empty entries. Example: ["Chinese", "English", "English"] -> "Chinese,English" Args: langs: List of canonical language names. Returns: str: Merged language string. NrrR)r1rr)routprevr:s r5merge_languagesrsVC D  W"OO   9  1  88C=r7)g@gY@rr<)5rTrburllib.requestr^ dataclassesrtypingrrrrrr urllib.parser renumpyrg soundfilerdr0rri AudioLike MaybeListrMAX_ASR_INPUT_SECONDSMAX_FORCE_ALIGN_INPUT_SECONDSrr*rrrr6r9r?rDrHrObytesrVrnrzrrrrrr~rrrrrr7r5rs  !>>!  "**c/   #tCy. !   #"T#Y@  )c)c), ` ` `191c1cd#$!S!U! ceBJJO4$=2::="**=   Y2::45U9d9o#=>54 CS5 %49%#%(492E%$ $, # V V VV V  V  % E! "# VrAL$(C# C#C=C# 38_C#L49r7