o \i&7@shUddlZddlZddlZddlmZddlmZmZm Z m Z m Z m Z ddl mZddlZddlZddlZe ee ejeffZe ee efZdZdZdZdZgd Ze eed <d Zd Z d edefddZ!d eddfddZ"dede efddZ#dede$fddZ%dede$fddZ&dede'fddZ(dede ejeffddZ)d ejdejfd!d"Z*d ejdejfd#d$Z+d%edejfd&d'Z,d(e ee efde ejfd)d*Z-d+e ed,edee efd-d.Z.ed/d0Gd1d2d2Z/ 3 4dFd5ejd6ed7e0d8e0d9e0de e eje0ff d:d;Z1dGd=d>Z2 dHd?ed@e ede eeffdAdBZ3dCe edefdDdEZ4dS)IN) dataclass)AnyIterableListOptionalTupleUnion)urlparsei>ig?)(ChineseEnglish CantoneseArabicGermanFrenchSpanish Portuguese IndonesianItalianKoreanRussianThai VietnameseJapaneseTurkishHindiMalayDutchSwedishDanishFinnishPolishCzechFilipinoPersianGreekRomanian Hungarian MacedonianAssameseBengaliGujaratiKannada MalayalamMarathiOdiaPunjabiTamilTeluguSUPPORTED_LANGUAGESz z language languagereturncCsH|durtdt|}|std|dd|ddS)a? Normalize language name to the canonical format used by Qwen3-ASR: first letter uppercase, the rest lowercase (e.g., 'cHINese' -> 'Chinese'). Args: language (str): Input language name. Returns: str: Normalized language name. Raises: ValueError: If language is empty. Nzlanguage is Nonezlanguage is empty) ValueErrorstrstripupperlower)r4sr=K/home/ubuntu/training/qwen3-asr-1.7b-phase2-sft/qwen_asr/inference/utils.pynormalize_language_nameTs   r?cCs |tvrtd|dtdS)z Validate the language is supported. Args: language (str): Canonical language name. Raises: ValueError: If unsupported. zUnsupported language: z . Supported: N)r3r7)r4r=r=r>validate_languagejs r@xcCst|tr|S|gSN) isinstancelist)rAr=r=r> ensure_listxsrEr<cCs4zt|}|jdvot|jWStyYdSw)N)httphttpsF)r schemeboolnetloc Exception)r<ur=r=r>is_url|s  rMcCs2|drdSd|vrd|vrt|dkrdSdS)Nz data:audioT/\F) startswithlen)r<r=r=r>is_probably_base64s rSb64cCs0d|vr|dr|ddd}t|S)N,zdata:r6)r9rQsplitbase64 b64decode)rTr=r=r>decode_base64_bytess rYcCst|r>tj| }|}Wdn1swYt|}tj|ddd\}}Wdn1s8wYn2t|rft |}t|}tj|ddd\}}Wdn1s`wYn t j |ddd\}}t j |t jd}t|}||fS)Nfloat32F)dtype always_2d)srmonor[)rMurllibrequesturlopenreadioBytesIOsfrSrYlibrosaloadnpasarrayrZint)rAresp audio_bytesfaudior]r=r=r>load_audio_anys$   rprocCsf|jdkr|S|jdkr+|jddkr |jd|jdkr |j}tj|ddtjStd|j)Nr6raxiszUnsupported audio ndim=)ndimshapeTrimeanastyperZr7)ror=r=r>to_monos  "r{cCs\|tj}|jdkr |Sttt|}|dkr|S|dkr%||}t|dd}|S)Nrg?g)rzrirZsizefloatmaxabsclip)ropeakr=r=r>float_range_normalizes  racCst|tr t|\}}n(t|tr+t|dkr+t|dtjr+|dt|d}}n tdt |t t |}|t krKt j||t dtj}t|}|S)a/ Normalize one audio input to mono 16k float32 waveform in [-1, 1]. Supported inputs: - str: local file path / https URL / base64 audio string - (np.ndarray, sr): waveform and sampling rate Returns: np.ndarray: Mono 16k float32 waveform in [-1, 1]. rqrr6zUnsupported audio input type: )orig_sr target_sr)rCr8rptuplerRrindarrayrk TypeErrortyper{rj SAMPLE_RATErgresamplerzrZr)rror]r=r=r>normalize_audio_inputs &raudioscCst|}dd|DS)NcSsg|]}t|qSr=)r).0rr=r=r> sz$normalize_audios..)rE)ritemsr=r=r>normalize_audiossrxs chunk_sizeccs@|dkr |VdStdt||D] }||||VqdS)z Yield chunks of a list. Args: xs (List[Any]): Input list. chunk_size (int): Chunk size. Yields: List[Any]: Slices of xs. rN)rangerR)rrir=r=r> chunk_lists rT)frozenc@s<eZdZUdZeed<eed<ejed<eed<eed<dS) AudioChunka[ One chunk cut from an original audio. Attributes: orig_index: Index of the original sample in the input batch. chunk_index: Index of this chunk within the original sample. wav: Mono float32 waveform. sr: Sampling rate. offset_sec: Start offset of this chunk in the original audio, in seconds. orig_index chunk_indexwavr] offset_secN) __name__ __module__ __qualname____doc__rk__annotations__rirr~r=r=r=r>rs    r@Y@rr] max_chunk_secsearch_expand_sec min_window_msc Cs6tj|tjd}|jdkrtj|ddtj}t|jd}|t|}||kr.|dfgSt||}t||}t dt|d|} g} d} d} || |kr| |} t | | |}t || |}||| krl| }n9|||}t |}tj |tj | tjdd d }tt|}|}|| }|||}tt|}|||}tt || d}tt ||}|| |}| || f| || t|7} |} || |ksQ|| |}| || ftt|}g}| D]+\}}|jd|kr|t|jd}tj|d|fd dd tj}|||fq|} | S) a Split a long audio into chunks close to max_chunk_sec, using a low-energy boundary. This implementation guarantees: - Concatenating all returned chunks reproduces the original audio exactly (total number of samples is identical, no overlaps, no gaps). Args: wav: Mono waveform float32. sr: Sampling rate. max_chunk_sec: Target max chunk duration in seconds. search_expand_sec: Boundary search half-window in seconds. min_window_ms: Sliding window in milliseconds for energy estimation. Returns: List[Tuple[np.ndarray, float]]: List of (chunk_wav, offset_sec). r_r6rsrtrr|g@@valid)modeconstant)rconstant_values)rirjrZrvryrzrkrwr~rminrconvolveonesargminappendMIN_ASR_INPUT_SECONDSpad) rr]rrr total_len total_secmax_lenexpandwinchunksstartrcutleftrightboundarysegseg_abs window_sumsmin_poswstartwendlocalinnerchunktailmin_lenpaddedcoffrr=r=r>split_audio_into_chunkss\               rcs2dd}dfdd |}|||}||}|S)NcSsg}d}t|}||krUd}|||kr2|||||kr2|d7}|||kr2|||||ks||krB|||||7}n|||||||7}||ks d|S)Nrr6)rRrjoin)r<threshresrncountr=r=r>fix_char_repeats[s    z4detect_and_fix_repetitions..fix_char_repeatsrcst|}|d}||kr|Sd}g}|||krd}td|dD]|}||||kr-nq||||} d} td|D]} || |} || | || krRd} nq<| r|} |||}|||kr||||| kr| d7} ||7}|||kr||||| ksm|| |||d|||}d}nq!|rn||||d7}|||ks|s|||dd|S)NrqrFr6Tr)rRrrr)r<rrrmin_repeat_charsrresultfoundkpatternrrep start_idx total_rep end_indexfix_pattern_repeatsr=r>rlsP         z7detect_and_fix_repetitions..fix_pattern_repeatsrr=)text thresholdrtext_rawr=rr>detect_and_fix_repetitionsZs ,  rraw user_languagec Cs|durdSt|}|sdSt|}|r||fS|}d}t|v}|r-|td\}}nd|fS|}d|vrG|}|sCdSd|fSd}|D]$} | } | sVqM| } | trq| t td} | rot | }nqM||fS)a 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)rrrr6z language none) r8r9r _ASR_TEXT_TAGrVr; splitlinesrQ _LANG_PREFIXrRr?) rrr< meta_part text_parthas_tag meta_lowertlanglinelowvalr=r=r>parse_asr_outputsB     rlangscCsFg}d}|D]}|p d}|sq||krq|||}qd|S)aM 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. NrrU)r9rr)routprevrAr=r=r>merge_languagess   r)rrrrB)5rWrdurllib.requestr` dataclassesrtypingrrrrrr urllib.parser rgnumpyri soundfilerfr8rrk AudioLike MaybeListrMAX_ASR_INPUT_SECONDSMAX_FORCE_ALIGN_INPUT_SECONDSrr3rrrr?r@rErIrMrSbytesrYrpr{rrrrrr~rrrrr=r=r=r>sz     + $" YF  F