(HJji dZddlZddlZddlZddlZddlZddlZddlZddlZddl m Z ddl m Z m Z mZmZmZmZmZmZddlmZddlmZddlmZddlmZmZmZmZm Z m!Z!m"Z"m#Z#edZ$dee$d e%d e$fd Z&gd Z'd e(d e)fdZ* dBde(de e(de)de e e(d e f dZ+dee(e fd e%fdZ,de d e%fdZ- dCdej\de%de%de e/d df dZ0de(de e(de(de%d ef d Z1 dDdee(e fde(de%d!e)d"e)d ej\f d#Z2da3da4da5da6da7d$Z8d%Z9d&Z:d'Z;d(Zd+e?fd,Z@ dFd-eejd.fd/e?d0e?d1e?fd2ZB dGd-eejd.fd3e=d4e?d5e?fd6ZC dHd-ee(ejfd0e?d+e e?d7e)d8e e?d ejf d9ZDgd:ZEd;e(d e)fd<ZFd=e(d e)fd>ZGde(de e(d e e(fd?ZHdee(e fd e e(fd@ZIde(fdAZJy)IzUtility functions for mlx_audio. This module provides a unified interface for loading TTS, STT, and STS models, with lazy imports to avoid loading unnecessary dependencies. N)Path)ListOptionalTupleTypeTypeVarUnion get_originget_type_hints)snapshot_download)STR_TO_WINDOW_FNbartlettblackmanhamminghanningistft mel_filtersstftT data_classdatareturnctj|st|dt|}i}tj|D]}|j }||vr||}||}t |}|tur.|jD cgc]} | tdus| } } | r| d}tj|rt|tr t||}|||<|di|Scc} w)Nz is not a dataclassr) dataclasses is_dataclass TypeErrorr fieldsnamer r __args__type isinstancedict from_dict) rr field_typeskwargsfield field_namevalue field_typeoriginaargss Y/Users/ahmed/devFolder/claude-voice/.venv/lib/python3.12/site-packages/mlx_audio/utils.pyr$r$-s  # #J /:,&9:;; ,K F##J/ZZ T ! Z  , J' U?)22J2!atDz6IA2DJ!!W   # #J /Jud4Kj%0E"z)0,   Ks C%C%) z*.json *.safetensorsz*.pyz*.modelz *.tiktokenz*.txtz*.jinjaz*.jsonlz*.yaml*.npzz*.pthpathc|jdxs>|jdxs+|jdxst|dkDxr|ddk(S)z5Check if the path looks like a local filesystem path../~:) startswithlen)r1s r._is_local_pathr:bsV  . ??3  . ??3  . IM ,d1gn path_or_hf_reporevisionforce_downloadallow_patternsct|j}|jr|St|rt d||t }tt ||||}|S)aP Ensures the model is available locally. If the path does not exist locally, it is downloaded from the Hugging Face Hub. Args: path_or_hf_repo: The local path or Hugging Face repository ID of the model. revision: A revision id which can be a branch name, a tag, or a commit hash. force_download: Force re-download even if cached. allow_patterns: File patterns to download. Defaults to common model files. Returns: Path: The path to the model. Raises: FileNotFoundError: If a local path is provided but doesn't exist. zLocal path not found: )r=r?r>)r expanduserexistsr:FileNotFoundErrorDEFAULT_ALLOW_PATTERNSr )r<r=r>r? model_paths r.get_model_pathrFlsz,o&113Jo&"88I JKK/ ))  J r;rEc t|tr t|fi|}|dz }|jr,t |d5}t j |cdddStd|#1swYxYw)a_Load model configuration from a path or Hugging Face repo. Args: model_path: Local path or Hugging Face repo ID to load config from **kwargs: Additional keyword arguments (revision, force_download) Returns: dict: Model configuration Raises: FileNotFoundError: If config.json is not found at the path z config.jsonzutf-8)encodingNzConfig not found at )r"strrFrBopenjsonloadrC)rEr& config_filefs r. load_configrOsp*c"#J9&9 },K + 0A99Q<1 0 2:,? @@1 0s A,,A5ctjt|dz }|s!tjt|dz }|std|i}|D]&}|jt j |(|S)zLoad model weights from safetensors or npz files. Args: model_path: Path to the model directory Returns: dict: Dictionary of weight name -> array Raises: FileNotFoundError: If no weight files found r/r0z.No weight files (safetensors or npz) found in )globrIrCupdatemxrL)rE weight_filesweightswfs r. load_weightsrWs}99So!=>?L yyZ'%9!:; .get_class_predicatesq.) 1h AHHNN2$6$Cq$H ,/15K+t,"" ? "G}''r;bitsmodeaffine)r^rjrkclass_predicate)getnnquantize)rXrYrUrZrir^r\s `` @@r.apply_quantizationrqs{::nd3Lzz"7> !!,3J((KK  & !   fh /+ r; model_type model_namecategorymodel_remappingc |j|d}ttj|z dz }g}|j rm|j r]|j D]J}|j s|jjdr0|j|jL|||k7r|D]}||vr|}||vs||}nn||} d|d|} tj| } | |fS#t$ri} | j k7r,tddtd |d | d | j| d |d |d} tj| t!| d} ~ wwxYw)a Retrieve the model architecture module based on the model type and name. Args: model_type: The type of model to load (e.g., "whisper", "voxtral"). model_name: List of model name components for remapping hints. category: Either "tts" or "stt". model_remapping: Dictionary mapping model names to architecture names. Returns: Tuple[module, str]: The imported architecture module and resolved model_type. Raises: ValueError: If the model type is not supported. Nmodels__ mlx_audio..models. T)flushz" Missing dependency while loading z: z& Please install it using: pip install z Model type z not supported for r3)rnr__file__parentrBis_diriterdirrr8append importlib import_module ImportErrorprintloggingerror ValueError) rrrsrtrumodel_type_mapped models_diravailable_modelsitempart module_patharchemsgs r.get_model_classrs,(++J=h&&1HF)r=r>r?Invalid model path type: rE architecturerllama acoustic_dimtada)rrrsrtru ModelConfigsanitizerZ)rpost_load_hook)popr"rIget_model_name_partsrnrFrrr!rOlowerrrdrr$ModelrWrgetattrrqlistitemsrSeval parametersr)rErtrurrr&rsrrr?r=r>rY model_class model_configrXrUrZs r.base_load_modelrAsD,.5JL$/JZZ 0$7N*c"  -j9J::j$/$4e<# ))  J %  -j9J4T*5E4FGHH  $Fz?F<ZZ d3 ZZ5 .8.DZ]((*$ W6!9 -' K ; . ))&1    l +E:&Guj!..)$E+BDIufg/DE tGMMO,V<    "# JJL{  "23!!00 C Lr;c*tddlm}|atS)zLazy load STT utils.rutils) _stt_utils mlx_audio.sttr) stt_utilss r._get_stt_utilsr4 r;c*tddlm}|atS)zLazy load TTS utils.rr) _tts_utils mlx_audio.ttsr) tts_utilss r._get_tts_utilsrrr;c*tddlm}|atS)zLazy load STS utils.rr) _sts_utils mlx_audio.stsr) sts_utilss r._get_sts_utilsrrr;c*tddlm}|atS)zLazy load VAD utils.rr) _vad_utils mlx_audio.vadr) vad_utilss r._get_vad_utilsrrr;c*tddlm}|atS)zLazy load LID utils.rr) _lid_utils mlx_audio.lidr) lid_utilss r._get_lid_utilsrrr;coeffcddl}|j|j|}|ddkrt|dd}||z dz}||dkD}|jd}|dkr|S|j |t d|zt d |z}||j||z dd z}|j|j|}|d kDr||z }|S) zNormalize the volume of an audio signal. Args: audio: Input audio signal array (numpy array). coeff: Target coefficient for normalization, default is 0.2. Returns: numpy array: The volume-normalized audio signal. rNrcg?gMbP?g{Gz? g?gGz?)a_mina_maxr6)numpysortabsmaxremeanintclip)audiornptempscaling_factorLvolume max_values r.audio_volume_normalizers 77266%= !D Bx#~T"Xt,&, t D 1 A Bw WWT#cAg,TAX7 8F BGGEFN#RG@ @Ervve}%I1} ! Lr;lengthc ddl}ddl}|jd|kr-|j|dt ||jdz f}|j d|jd|z }t ||z}|||S)zGet a random audio segment of given length. Args: audio: Input audio array (numpy array). length: Desired segment length (sample_rate * duration). Returns: numpy array: Audio segment of specified length. rN)randomrrepadrrandint)rrrr start_index end_indexs r.random_select_audio_segmentrsz {{1~uq#fu{{1~&=">?@..EKKNV$;?? >>"**5> 11+ s 6Fvolume_normalizesegment_durationc t|tjr|St|tst dt |ddl}ddl}ddlm }|jj|std|||\}} |j} t| dkDr|jd}|| dz }|| k7r t!|| |}|t#||z} t%|| }|r t'|}|E|jd|kDr|d|}n-|j)|dt#||jdz f}tj|tj*S) aLoad audio from file path or return mx.array as-is. Args: audio: Audio input - can be: - str: Path to audio file (will be loaded and resampled to sample_rate) - mx.array: MLX array (returned as-is) sample_rate: Target sample rate (default 24000) length: Target length in samples (pad or truncate if specified) volume_normalize: Whether to normalize audio volume segment_duration: If specified, randomly select a segment of this duration (seconds) Returns: mx.array: Audio waveform at target sample rate Raises: FileNotFoundError: If audio file path does not exist TypeError: If audio is not str or mx.array z#audio must be str or mx.array, got rN)readzAudio file not found: r6)r)dtype)r"rSrrIrr!osrmlx_audio.audio_iorr1rBrCrer9sumrrrrrr) rrrrrrr audio_readsamplesrre seg_lengths r. load_audior$sW2%" eS !=d5k]KLL 5 77>>% "8 @AA *5 1G  MME 5zA~++1+%E!H$&& *:KH#'778 -gzB(1 == f $gv&GffWq#fw}}Q7G.G*H&IJG 88G2:: ..r;)rrrrr rrrr$rrrrr$is_valid_module_nameget_model_categoryr load_modelrFrOrWrqrrrcvt|txr(|jxrtj| S)z0Check if a string is a valid Python module name.)r"rI isidentifierkeyword iskeyword)rs r.r%r%s1 dC XT%6%6%8 XARARSWAX=XXr;rc tjj|duS#t$r4}|jxsd}||k(s|j |drYd}~yd}~wwxYw)Nr3F)rutil find_specModuleNotFoundErrorrr8)rexc missing_names r._has_model_moduler3sa~~'' 4D@@ xx~2 , &+*@*@L>QRAS*T s # A )AAA ct}t}t}t}t }|g|xsgz}d|j fd|j fd|j fd|j fd|j fg}|D]H\} } | |vr |D]9} | j | | } t| s!d| d| } t| s5| ccSJ|D];\} } |D]1} | | vs| | } t| sd| d| } t| s-| ccS=|D]6\} } |D],} | | vst| sd| d| } t| s(| ccS8y) zMDetermine whether a model belongs to the TTS, STT, STS, LID, or VAD category.ttssttstslidvadryrzN) rrrrrMODEL_REMAPPINGrnr%r3)rrrsrrrrr candidates categoriesrtremaphintrrs r.r&r&s I I I I I!1r2J  ))*  ))*  ))*  ))*  ))* J&% : % D99T4(D'-&xj?K - &&%Du}T{+D1 *8*HTFC $[1#O&&%D5 %9$%? *8*HTFC $[1#O & r;cd}t|tr#|jjdd}nvt|trO |j j d}|j |dzjjdd}ntdt|g}t}|jdDcgc]}|s| }}|D]}|r||vr |j||j|d|vr@|jdD],}|s||vs |j||j|.tjd d |}|s||vs|j||j|tt!|D]x} t| d zt!|dzD]X} || | } dj#| d j#| fD],} | s| |vs |j| |j| .Zz|S#t$r|jj}YwxYwcc}w) Nr4rchubr6z--r-_z [^a-z0-9]+r-r)r"rIrsplitrpartsindexrrr!setraddresubrr9join) rErsrErDseenr dash_partssubpart normalizedstartendsegmentcombineds r.rr,sJ*c"%%'--c226 J % 1$$**51E#))%!)4::<BB4HLJ4T*5E4FGHH E 5D#-#3#3C#8A#84D$#8JAtt|  T  $;::c?wd2LL)HHW%+ VVM2t4 *D0 LL $ HHZ "s:'C Oa$78C s+G XXg.0@A 4LL*HHX&B9( LG 1#..0J 1BsA H I I #IIc*t|}d} t|}|jdd}||jdd}|t j |}t ||}|s||td|tjtjt jtjtjd}||vrtd|d|||S#t$r}i}|}Yd}~d}~wwxYw)a$Load a TTS, STT, STS, LID, or VAD model based on its configuration and name. Args: model_name (str): Name or path of the model to load Returns: The loaded model instance Raises: ValueError: If the model type cannot be determined or is not supported Nrrrz#Could not determine model type for )r5r6r7r8r9z Model type 'z' not supported) rrOrCrnrinfer_model_type_from_configr&rrr'rrr)rsr load_errorrYr1rrmodel_category model_loaderss r.r'r'Zs%,J7JZ( L$/JZZ5 #%BB6J ' 4DEN   ! >zlKLL********** M]*<'7GHH (= ( 449  s C;; DD  D)NFN)N)FF)g?)rc)rii)i]NFN)K__doc__rrQrimportlib.utilrKr*rrHpathlibrtypingrrrrrr r r mlx.corecorerSmlx.nnrohuggingface_hubr mlx_audio.dspr rrrrrrrrr#r$rDrIboolr:rFrOrWModulecallablerqrrrrrrrrrrrrfloatrrrrrrr$__all__r%r3r&rr'rr;r.rfs       -    CL $q'  ! L # *. ,,sm,,T#Y' ,  ,^AE#t),A4A0TdD15 / 99/ //$H- /  /d;;S ;; ;  ;D _c4i ___  _  _YY_F     ''T(s(4 : <' (:::  :~ $2 <' ($2 $2$2 $2R "&* D/ bhh D/D/ SMD/ D/ sm D/ XX D/N @YsYtY 34232DI2(3-2j+U39%5+$s)+\-53-5r;