(HJj)! UdZddlZddlZddlZddlZddlZddlZddlZddlZddl Z ddl Z ddl Z ddl m Z ddlmZmZmZddlmZddlmZmZmZmZddlmZddlmZddlZddl Z ddl!Z!ddl"m#Z#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z*m+Z+dd l,m-Z-dd l.m/Z/m0Z0m1Z1dd l2m3Z3dd l4m5Z5dd l6m7Z8ddl6m9Z:ddl;mZ>m?Z?m@Z@mAZAddlBmCZCmDZDmEZEmFZFmGZGddlHmIZImJZJddlKmLZLdedefdZMGddZNe#ZOde#deePfdZQejdZSeSr-eSjdDcgc]}|jc}ndgZVeQeOeVe de#fdZWeWeOj_YGdd e5ZZGd!d"e5Z[Gd#d$e5Z\eNZ]ejZ_da`eeDead%<eGd&d'ZbeGd(d)ZceGd*d+Zdd,ePfd-Zed,ePddfd.Zfd/d0hZgGd1d2eCZhGd3d4ZiGd5d6eCZjGd7d8eCZkdeDfd9Zld:eEdeGfd;Zmd:eEdZoeOjd?d@ZqeOjdAdBZreOjdAd,ePfdCZteOjdAd,ePfdDZveOjdEdFeZd@chunk_durationframe_thresholdrcontextprefill_step_sizetextword_timestampstimestamp_granularities)rRrSrTrUrrrrVrrrr5rrrrrrrr rWrIr=rrs JHcDjGTJ NE OSFDGS4Z!s!D#*!OT!-1Xc]1rIrc,eZdZUeed<eed<eed<y)SeparationResponsetargetresidual sample_rateN)rRrSrTrUrrrWrIr=rrs KMrIrrhceZdZUeed<y)SpeechTaskPayloadrequestN)rRrSrTrlrrWrIr=rrs rIrcJeZdZUeed<eed<ejed<eed<y)TranscriptionTaskPayloadrfilenameaudiorN) rRrSrTrrrUr6ndarrayrrWrIr=rrs !!M ::rIrc^eZdZUeed<ej ed<eed<eed<eed<eed<y)SeparationTaskPayloadrJrr descriptionmethodstepsN)rRrSrTrUrr6rrrWrIr=rrs'O :: K JrIrrJc,tj|SrAmodel_providerr+rJs r=_load_model_for_inferencers  $ $Z 00rIcK tjt|d{y7#t$rt$r}tdd|d|d}~wt $r}tdd|d||d}~wwxYww) aLoad ``model_name`` synchronously and translate failures into HTTPException. Routes that return a ``StreamingResponse`` commit the HTTP status + headers before the body generator runs, so any failure inside the inference worker surfaces as 200 OK with an empty body. Pre-flighting the load here lets the framework's exception handlers turn the failure into a real HTTP error response. Warm models are a no-op (``ModelProvider.load_model`` is cached). NzModel not found: z. is not a known HuggingFace repo or local path status_codedetailizFailed to load model z: )rC to_threadrrr Exception)rJexcs r=_preflight_model_loadrs  9:FFF  "&zn4bc   *:.3%@  s<A5)')A5)A2A  A2A--A22A5rrceZdZdeddfdZy)STTExecutionAdapterrr-Nc |j}tjj|jxsd\}}|xsd}dt j d|j |}t||j|j t|j}|jjdhd}tj|j } |j#D cic]\} } | | j$vs| t&vr| |  }} } |j |fi|} t)| drt)| d rd} | D]}|j*j-rnt/|t0r | |z } || d }n>|j2t5|d dt5|d dt5|d dt5|ddd}|j7t9j:t=|dznJ|j*j-s0|j7t9j:t=| dztjj?|rtj@| |jCycc} } w#tjj?|rtj@|wwxYw)N.mp3z/tmp/_rmTexclude exclude_none__iter____next__)r accumulated start_timeend_timeis_finalr)rstartendrr )"payloadospathsplitextrtime request_id audio_writerrrrJr model_dumpinspect signaturegenerater2 parameters_STT_EXTRA_KWARGShasattr cancel_eventis_setr/rUrgetattr emit_datajsondumpsr3existsremove emit_done)rGrrrextsuffixtmp_path stt_model gen_kwargsrkeyvalueresultaccumulated_textchunk chunk_datas r= run_serialzSTTExecutionAdapter.run_serialsg,3OO!!'"2"2"8b93499;-q););(v-F"G$"NOww~~h' (#C<ww~~h' (#(s A%J0#JD'JJ7J=)rRrSrTr'rrWrIr=rrs."2.t.rIrc^eZdZ d dZedefdZdeddfdZd dZ de ddfd Z d d Z y) _TTSAdapterContinuousSessioncZ||_||_i|_t|_d|_y)Nr)adapter model_session _requestsset_emitted_audio_next_sequence_id)rGrrs r=rHz%_TTSAdapterContinuousSession.__init__Bs,  *68(+!"rIr-cL|jjxr |j SrA)ridlerrFs r=rz!_TTSAdapterContinuousSession.idleNs !!&&=t~~+==rIrNc |jjr|jy|j}|j}|j }|xj dz c_||j |<|jjt||j|j|j|j|j|j|j |j" gy)N) sequence_idrrprorrrtrurxry)rrrrrrrraddr)rnrprorrrtrurxry)rGrrspeech_requestrs r=submitz#_TTSAdapterContinuousSession.submitRs    & & (     %,__ ,,  !#&-{#  +'--(..+44(..)00(..,66+44   rIc|j|jj}|D]V}|jj |j }|,|j w|j|j |j|jj|j d|jj|j |j|jdkDr|jj}|j w|jt#d|j|jj|j d|jj|j c|j$j'|||j|j |jj)|j |j*s|j |jvr|jt-dd|j|jj|j d|jj|j Yy)Nrz2TTS batch event included audio without sample_rateNo audio generatedr)_cancel_disconnected_requestsrsteprgetrerror emit_errorrpoprdiscardrsamplesrrr ValueErrorr _emit_audiordoner)rGeventseventrrs r=rz!_TTSAdapterContinuousSession.stepls **,##((*Enn(():):;G{{&""5;;/!!#""5#4#4d;##++E,=,=>{{&5==1+<070G0G$$,&&"#WX%%'NN&&u'8'8$?''//0A0AB (("KK%%  ##''(9(9:zz$$D,?,??&&%#>RS!!#""5#4#4d;##++E,=,=>KrIrct|jjD]#}|j||j %|jj |j j yrA)r4rvaluesrrclearr)rGrrs r=failz!_TTSAdapterContinuousSession.fails\DNN1134G   u %    5  !!#rIcTt|jjD]\}}|jj s!|j j ||j|jj|d|jj|yrA) r4rr2rrrcancelrrrr)rGrrs r=rz:_TTSAdapterContinuousSession._cancel_disconnected_requestss$()=)=)?$@ K''..0    % %k 2     NN  {D 1    ' ' 4 %ArI)rTTSExecutionAdapterr-N) rRrSrTrHpropertyrVrr'rr BaseExceptionrrrWrIr=rrAs^ #' #>d>> . 4 4(?T$-$D$5rIrceZdZdZdZdefdZdedefdZdede e e ffdZ de fd Z ded edefd Zdedefd Zdedefd ZdedefdZdedefdZdede fdZdede fdZdefdZdedededdfdZdeddfdZdeeddfdZy)r_mlx_audio_loaded_modelctjdd} tdt||_y#t $r d|_YywxYw)NMLX_AUDIO_TTS_MAX_BATCH_SIZE8r)rgetenvmaxrmax_batch_sizer)rGraw_max_batch_sizes r=rHzTTSExecutionAdapter.__init__sEYY'EsK $"%a-?)@"AD  $"#D  $s3AArct||jd}|,t|j}t ||j||SrA)r_REQUEST_MODEL_ATTRrrJsetattrrGrrms r=_get_model_for_requestz*TTSExecutionAdapter._get_model_for_requestsB!9!94@ =-g.@.@AE GT55u = rIr-c  |jj}t|j|j|j |j |j|j|j|j|j|j S)N) rzr{r}r~rrwrrrr) rrr*rzr{r}r~rrwrrrrrGrrs r=_build_batch_optionsz(TTSExecutionAdapter._build_batch_optionssx(/(?(?&22 && &&-@@%00$..!((-@@.."**  rIcV|jj}|jdhdS)NrmTr)rrrr%s r=_request_kwargsz#TTSExecutionAdapter._request_kwargss)(/(?(?(('(NNrInamect||d}t|syt|jdk(r|t |vry|S)Nz unittest.mock)rcallabler0rSvars)rGrmr)rs r=_get_callable_model_attrz,TTSExecutionAdapter._get_callable_model_attrsAtT* ; ! !_ 4T%[9P rI continuousc \|jj}|rdnd}|j||}t|r t |di|j |S|j ry|js |jry|jdvry|jdvry|jdvryy)Nsupports_tts_continuous_batchsupports_tts_batchFNrsNrqTrW) rrr-r+rVr(rrxryrtrrru)rGrmrr.r hook_namehooks r=_model_supports_requestz+TTSExecutionAdapter._model_supports_requests)0(?(?/9 +?S ,,UI> D>=t33G<=> >    # #~'>'>   6   { 2   { 2rIc|j|d}|y|j||dsy|jj}|jryt j |}|j}d|vry|jrd|vry|jrd|vry|jdvrd|vry|jd vrd |vry|jd vrd |vry|jrd |vry|jrd |vryy)Nbatch_generateFr.textsvoices instructsr2gendersr3speedspitches ref_audios ref_textsT)r-r6rrrrrrrprortrrrurxry)rGrmrr8rrparamss r=_can_call_batch_generatez,TTSExecutionAdapter._can_call_batch_generates66u>NO  !++E7u+M(/(?(?  %%n5 %% &    HF$:  " "{&'@   69F;R   { 2xv7M   { 2y7N  # # F(B  " "{&'@rIcZ|j|d}|duxr|j||dS)Ncreate_tts_batch_sessionTr9)r-r6)rGrmrcreate_sessions r=_can_call_continuous_sessionz0TTSExecutionAdapter._can_call_continuous_sessionsF66  & T) d.J.J 7t/K/  rIcH|j|}|j||SrA)r#rCr"s r=supports_batchz"TTSExecutionAdapter.supports_batchs%++G4,,UG<fi| D]} |j@jCr&tEjF|jIy|j4r)|jK||| jL| jx|jO| jL| | j} |j4r|jIy|s tddtQjR|} |jK||| | |jIy)Nr Reference audio file not found: rr) load_audio model_typespark)rvolume_normalizerprrrtrurorwrxryrzr{r}r~rrrrr)*rrr#rxr/rUrrrrmlx_audio.tts.generater]rr^rrprrrtrurorwryrzr{r}r~rrrrrrnrrmx clear_cacherrrappendr6 concatenate) rGrrrrmrxr] normalize audio_chunksrgenerate_kwargsrconcatenated_audios r=rzTTSExecutionAdapter.run_serialSs%,__ ++G4",, Is377>>),# #=i[I :|4T9I9IW9TI"!--!*I    ^)) ^))  n++  ^))  //  11     //  >55  ^))  ^))  !."C"C  n++  !."C"C  .33 ~--! &%enn^%9%9M_MF##**, !!#$$  "LL&& ##FLL1&"("4"4K!N$      C8LM M^^L9 .2DkRrIrequestsct|dk(rj|dyj|dtfd|Ds|D]}j|y|djj }|Dcgc]"}|jj j $}}|Dcgc]"}|jj j$}}|Dcgc]"}|jj j$}}|Dcgc]"}|jj j$}}|Dcgc]"}|jj j$}}|Dcgc]"}|jj j$} }|Dcgc]"}|jj j$} }|Dcgc]"}|jj j$} }jd} | |D]}j|ytj | } id|d|d|d|d |d | d | d | d |j"d|j$d|j&d|j(d|j*d|j,ddd|j.}|j1Dcic]\}}|| j2vs||}}}|Dcgc]}g}}|Dcgc]}d}}| di|D]o}|j4}|dks|t|k\r#||}|j6j9rC||j;|j<|j>||<qtA|D]\}}|j6j9r|jC1||}||}|r|-|jEtGdd|jClt|dk(r|dntIjJ|}jM||jj |||jCycc}wcc}wcc}wcc}wcc}wcc}wcc}wcc}wcc}}wcc}wcc}w)Nrrc3BK|]}j|ywrA)rC).0rrmrGs r= z0TTSExecutionAdapter.run_batch..s" IQgD ) )% 9sr8r:r;r<r>r=r?r@rArzrwrr}r{r~rFrrrrrW)'lenrr#allrrrnrprorrrtrurxryr-rrrzrwrr}r{r~rr2r sequence_idxrrrdrr enumeraterrrr6rer)rGrjrfirst_speech_requestr:r;r<r>r=r?r@rAr8rkwargsrrraudio_chunks_by_sequence sample_ratesrrqchunksrrrms` @r= run_batchzTTSExecutionAdapter.run_batchs x=A  OOHQK ( ++HQK8 IQ  $($ .6qk.A.A.I.I>FGh7((..hG?GHxG'//))//xHEMNX'W__,,55X N?GHxG'//))//xHAIJg7??**11J@HIW7??**00IGOPxGgoo--77x PEMNX'W__,,55X N66u>NO  !#($ %%n5  U f    f  w  w   *    /;;  -77  .99  )//  )//  !"6"I"I  e +33! &*0 )7:33)BVBV;VCJ  BJ4JAR 4J8@)A1$ )A$.v.F!..La<3x=#@|,G##**, $\ 2 9 9&,, G)/););L &/&/x%8 !L'##**,!!#-l;F&|4K[0""!c:NO!!#!$V!1F1Ir~~f7ME   ''      +&9gHHNHJIPN6 5K)AsH'Q2'Q 'Q 'Q9'Q&'Q 'Q%'Q*6Q/ Q/ Q5( Q:)rRrSrTr rHr'r#r*r&r1rUrr(r-rVr6rCrGrIrKrNrPrSrlrrrr4rxrWrIr=rrsd3$.>  ,<    O'7ODcNO C "    67GD> ;K PT =&6=4=A1AAdA !1 c "',<''  7G  -!-&-  -  - F"2FtFPV $'7"8V TV rIrc.eZdZdZdefdZdeddfdZy)SeparationExecutionAdapterc.d|_d|_d|_yrA) _model_name_model _processorrFs r=rHz#SeparationExecutionAdapter.__init__s'+ rIrJc|j|k(r$|j|j|jfSddlm}m}|j ||_|j ||_||_|j|jfS)Nr)SAMAudioSAMAudioProcessor)r|r}r~ mlx_audio.stsrrfrom_pretrained)rGrJrrs r=_get_resourcesz)SeparationExecutionAdapter._get_resourcessn   z )dkk.E;;/ /=+;;JG..z: %{{DOO++rIrr-Nc |j}dtjd|jd}t||j|j  |j |j\}}||jg|g}d|jdzz }|j|d}|j|j|j|j|j|d} t!j"t%j&| j(d } t%j&| j*d } |j } d } |j-t/| | | | | | | t0j2j5|rt1j6| |j9y#t0j2j5|rt1j6|wwxYw) Nz/tmp/separation_rz.wav) descriptionsaudios)r step_size2)rr anchor_idsanchor_alignmentode_optode_decode_chunk_sizerctj}t|||d|jdt j |j jdS)NwavrUrutf-8)rWrXrseekbase64 b64encoderdecode) audio_arraysrrZs r=audio_to_base64z>SeparationExecutionAdapter.run_serial..audio_to_base64#sHFKEB A'' 6==gFFrI)rrr)rrrrrrrrJrrr separate_longrrrrrbrcr6arrayrrrrrrrrr)rGrrrrm processorbatchrrr target_audioresidual_audiorrs r=rz%SeparationExecutionAdapter.run_serials)0%diik]!G4F4F3GtLHgmmW-@-@A( $#2273E3EF E9%112 zE W]]Q./I!(iHG((||"// ++!&!7!7&( )F NN 88FMM!$45LXXfooa&89N++K G   "*<E,^[I + ww~~h' (#ww~~h' (#(s D&G7G7)rRrSrTrHrUrr'rrWrIr=rzrzs' , ,/"2/t/rIrzctZt}|jdt|jdt |jdt |atS)NsttrM separation)rhr%register_adapterrrrz)brokers r=get_inference_brokerr7sX "':'<=':'<= .H.JK! rIhandlecpKtj|jjd{S7wrA)rCr result_queuer)rs r=_next_inference_chunkrBs)""6#6#6#:#:; ;; ;s -646rcK t|d{}|jdk(rnq|jdk(r |j|jt j dd{|j d{r|jn|jy77A7+#|jwxYww)Nrrr)rkindrrrCsleepis_disconnectedr)rrrs r=_stream_inference_resultsrFs /77EzzV#zzW$kk!-- --" " ",,...   8 #.  sQCB0B*AB0*B,+B0B.B0C*B0,B0.B00CCcKd} t|d{}|jdk(r||jS|jdk(r |j|j}\7M#|jwxYww)Nrr)rrrrr)rrrs r=_await_inference_resultrWso F /77EzzV#   zzW$kk!]]F 7  s*A:A%A#A%A:)A%%A77A:/cKddiSw)NmessagezMWelcome to the MLX Audio API server! Visit https://localhost:3000 for the UI.rWrWrIr=rootres b sz /v1/modelsc Ktjd{}g}|D]3}|j|dtt jdd5d|dS7Cw)zG Get list of models - provided in OpenAI API compliant format. Nrmsystem)idobjectcreatedowned_byr4)rdata)rrQrdrr)rB models_datarms r= list_modelsrlsd "668 8FK!tyy{+$   k 229sA AAA cFKtj|dd|ddSw)z Add a new model to the API. Args: model_name (str): The name of the model to add. Returns: dict (dict): A dictionary containing the status of the operation. successzModel z added successfullystatusrrrs r= add_modelrs+j)fZL@S,T UUs!cKt|jd}tj|d{}|r t dSt dd|d7#w) a Remove a model from the API. Args: model_name (str): The name of the model to remove. Returns: Response (str): A 204 No Content response if successful. Raises: HTTPException (str): If the model is not found. "N)rrzModel 'z ' not foundr)r striprrNrr)rJremoveds r=rNrNsY$**3/J"// ;;GC((gj\4UVV Generate speech audio following the OpenAI text-to-speech API.rr\rNrM)rrmTr endpoint_kindrJrnormalized_kwargsrzaudio/zContent-Dispositionzattachment; filename=speech.) media_typeheaders)rxr/rUrrrrrrmrrrrrrrr)rrrs r= tts_speechrsZ(9(93?ww~~g//09':K:K9LM   ... ! # * *==!'2!,,gYT,R~~ +F !&'2G3345 !%A'BYBYAZ#[  /sBD D BD z/v1/audio/voicesrmcZK|s tdd ddlm}||dg}t |d z }|j sd|gd Std |jdD}d||Dcgc]}||d c}d S#t$r}d|gt |d cYd }~Sd }~wwxYwcc}ww)auList available voices for a TTS model. Resolves the model's HF snapshot directory and enumerates ``voices/*.safetensors`` files (the convention used by Kokoro and similar voice-pack-based TTS models). Returns an empty ``data`` list for models that don't ship per-voice packs so callers can fall back to whatever defaults make sense for that model. rz!model query parameter is requiredrr)snapshot_downloadzvoices/*.safetensors)repo_idallow_patternsr4)rrmrrNr;)rrmrc34K|]}|jywrA)stem)rmps r=rnztts_voices..sE$DqAFF$Dsz *.safetensors)rr)) rhuggingface_hubrrrUris_dirsortedglob)rmrsnapshote voices_dirr;r;s r= tts_voicesrs 4WXXO5$+A*B  h(*J     5"== EJOOO$DE EF/56v!1%v6  O 5"s1vNNO7s@B+BA B+/ B&;B+ B# BB#B+B##B+z/v1/audio/transcriptions.FrrrrndjsonfilerrrrrrrrrrcrKt|||||||| | | | | }|jd{}tj|}t |d\}}|j t |jd{tjd|jt||jxsd|||jdhd |j }| d vrd}d } t|d{}|jdk(rn|jdk(r |j |j"}t%|t&r|j)dd}t+|j-D]t}|j/}|s t1j2|}t%|t6r d|vsd|vr|}Gt%|t6sXd|vs]||j9dxsd z }v |j;|d|i}| dk(r,t=|j9dxsd j/S| dk(r.t?d|j9dxsd j/iSt?|StAtC||dS7\7 7#t4$rY*wxYw#|j;wxYww)a6Transcribe audio using an STT model. The default ``response_format`` (``ndjson``) preserves mlx-audio's native ``application/x-ndjson`` streaming transport, where each line is a JSON object emitted by the underlying STT model (text deltas while streaming, or the full whisper segment payload for batch transcription). For OpenAI Audio API compatibility, ``response_format`` also accepts: * ``text`` -- ``text/plain`` body with the final transcript only. * ``json`` -- ``application/json`` body shaped ``{"text": "..."}``. * ``verbose_json`` -- ``application/json`` body with the full payload from the underlying model (``text``, ``segments``, ``language``, ...), passed through unchanged. See https://platform.openai.com/docs/api-reference/audio/createTranscription ) rmrrrrrrrrrrrNF always_2drz audio.mp3)rrrrrmTrr)rr verbose_jsonrrrrreplace)errorssegmentsrrrzapplication/x-ndjson)r)"rrrWrX audio_readcloserrmrrrrrrrrrrr/bytesrrU splitlinesrrloadsrr1rrrrrr)rrrmrrrrrrrrrrrrrrtmprrrfullrrlinerawr,s r=stt_transcriptionsrsF#%'+ ' 7 G D **T C3%0IE2IIK  ... ! # * *==(]]1k  ",,gYT,R~~+ F::#  3F;;::'::(++%}}dE*;;wy;ADt9//1C))+C !"jjo"#t,"c)Z3->"#C.6S=#swwv'<"< 2* MMO <K(D f $$dhhv&6&<"%C%C%EF F f $$((6*:*@b)G)G)I JK KD!! !&'2) y  /(<%! ! MMOs+J7J AJ7?J A,J7-J"<J=B J" J +J" J"J"-BJ7 J7J" JJ"JJ""J44J7z/v1/audio/separationsz"mlx-community/sam-audio-large-fp16speechmidpointrrrc 2K|jd{}tj|}t|d\}}|j t j d|t|||||||||d} t| d{S7|7w)aSeparate audio using SAM Audio model. Args: file: Audio file to process model: SAM Audio model name (default: mlx-community/sam-audio-large-fp16) description: Text description of what to separate (e.g., "speech", "guitar", "drums") method: ODE solver method - "midpoint" or "euler" (default: midpoint) steps: Number of ODE steps - 2, 4, 8, 16, or 32 (default: 16) Returns: JSON with base64-encoded target and residual audio, plus sample rate NFrr)rJrrrrr)rrr)rrJrr) rrWrXrrrrrr) rrmrrrrrrrrs r=audio_separationsrUs( D **T C3%0IE2IIK ! # * *"%#  ' +F")0 00- , 1s"BBA7BBBB websocketrr is_partial streamingc Kdtj|jjv}|r|r|jt j |d|d}d} |} |D]^} t | tr| nt| dt| } | | z } t| dd} | r| | } |jd | d d{`|jd | d| |d d{yd tjd}t||| |j||d}t|dr!|jrt|jnd}|j|j|t|d||dd{t j"j%|rt!j&|yy777>#t j"j%|rt!j&|wwxYww)aHandle both streaming and non-streaming model inference over WebSocket. Streaming models (whose generate() accepts a ``stream`` parameter) receive the audio as an ``mx.array`` and yield token deltas sent as ``{"type": "delta", "delta": "..."}`` messages, followed by a ``{"type": "complete", ...}`` message. Non-streaming models fall back to temp-file + batch generate, sending the legacy ``{"text": ..., "is_partial": ...}`` format. rTF)rrrrrrNdelta)r0rcomplete)r0rrrrz/tmp/realtime_r)rrr)rrrr)rrrrrbrr/rUr send_jsonrrrrr3rrrrr)rrrrrrrsupports_stream result_iterrdetected_languagerr chunk_langrrrs r=_stream_transcriptionrs&'"3"3I4F4F"G"R"RRO9(( HH[ !$5)  $ E#E3/WUFCPUJ5W  5 K  D9J/7$.!%%w&GH H H!!!"# -(     $DIIK=5Hk;7 $''8U'SF6:.6??"&//2  %%"KK ( ' H E",   ww~~h' (#(= I & ww~~h' (#(sOB8G3:F3; G3F5*G3A1F98F79F9=7G35G37F997G00G3z!/v1/audio/transcriptions/realtimec K|jd{ |jd{}|jdd}|jdd}|jdd}|jdd}td |d |d |d |td tj |}tdt jd}d}t||zdz } td| d|d|dg} t|dz} t|dz} t|dz} d}d}tj}d}d}|jdddd{td |jd{}d|vrtj|dtj}t!|| z}d}d}t#|D]H}|| z}|| z}|||}t!|| k(s! |j%|j'|rd}|d z }Jtj}|r|j-tj.d"z }| j1||d z }d}|}t!| |d#zzt!|krKtd$|d%|d&t!| dt!| |z d'd( n|d z }|d)zdk(rtd*|d+||z }d} d}!t!| dkDr|s6t!| | k\r(|r&d} td,| |z d'd-t!| |z d'd.nb||k\r1t!| | k\r#d}!td/|d'd0t!| |z d'd.n,t!| | k\rd}!td1t!| |z d'd.| rEt!| | k\r7| }"tj2| d|"}#d} t5|||#||d|2d{|!rt!| dkDrt!| }"tj2| }# t5|||#||d|2d{g} d}td6|"d|"|z d'd7n3d9|vr/ t=j>|d9}&|&jd:d;k(rn  |jCd{y7j7T7,7 #t$rY1wxYw#t(t*f$r }td!|d}|d z }Yd}~d}~wwxYw7#t$rU}ddl}$t9|}%|$j;td3|%|j|%d4d5d{7Yd}~^d}~wwxYw7 #t$rU}ddl}$t9|}%|$j;td8|%|j|%d4d5d{7Yd}~)d}~wwxYw#t$rY=wxYw#t@$rYKt$rB} |jt9|d4d5d{7n#t$rYnwxYwYd}~d}~wwxYw7#t$rYywxYw# |jCd{7w#t$rYwwxYwxYww)rTzConfiguration received: model=z , language=z, sample_rate=z , streaming=zLoading STT model...zSTT model loaded successfullyizVAD initialized: frame_size=z samples (zms at zHz)g?g?g@rFreadyzReady to transcriberrdtyperzVAD error (assuming speech): @rzSpeech detected (rz frames): buffer z.2fzs)zSilence detected: skipped z chunksz1Processing initial chunk for real-time feedback: zs, total buffer: szProcessing due to silence gap: zs silence, buffer: z#Processing due to max buffer size: )rrz$Error during initial transcription: r)rrzProcessed final chunk: zs), buffer clearedzError during transcription: ractionstop)"accept receive_jsonrprintrr+ webrtcvadVadrrrreceiverr6 frombufferint16rorange is_speechtobytesrOSErrorastypefloat32extendrr tracebackrU print_excrrrr)'rconfigrJrrrrvadvad_frame_duration_msvad_frame_size audio_buffermin_chunk_sizeinitial_chunk_sizemax_chunk_sizesilence_skip_countspeech_chunk_countlast_speech_timesilence_threshold_secondsinitial_chunk_processedraudio_chunk_int16 num_frames has_speech speech_framesi frame_start frame_endframer current_timeaudio_chunk_floattime_since_last_speechshould_process_initialshould_process_final process_sizerr error_msgrs' r=stt_realtime_transcriptionsr9s    ` --//ZZ D ::j$/jj6 JJ{D1  ,ZL H:^\g[hhtu~t A $%"--j9  -.mm  !#[+@@4GH *>*:*EZD[[abmannq r  [3./ #   #  99;$'!"'!!WAV"WXXX #$  ) 1 1 33'!$&MM''2B"(($S! !!23~E " ! z*A"#n"4K +n TUX=YYlmpq}m~@KnKLOmPPQR\*nRX>V%RD*V%RV%,R ?R R A*V%0'RFV%1SS S 0V%<T4T1T40V%6,V"V%'X:W>;X?X>V%V% R RV%RV%S *S?V%S  V% S T.AT)T T)#V%)T..V%1T44 V=AV VV V% VV% V"V%!V""V%% W;.X1W;: W!WW! W6! W-*W6,W--W60X6W;;X>X X  X> X  X>X;X,%X(&X,+X;, X85X;7X88X;;X>i]chtjd}|sy t|S#t$rYywxYw)N)MLX_AUDIO_REALTIME_TRANSCRIPTION_DELAY_MS)rrrr)rs r=_default_transcription_delay_msr<s7 ))? @C 3x s % 11rzdelay_mscd|i}|2tj|j}d|jvr||d<|jdi|S)zsOpen a streaming session, forwarding ``transcription_delay_ms`` only to models that declare the parameter. rztranscription_delay_msrW)rrcreate_streaming_sessionr)rmrzr=rtsigs r=_open_streaming_sessionrBsW";/F > >? #s~~ 5/7F+ , )5 ) ) 3F 33rIpcm16 from_rateto_ratec(|jtjdz }||k(s|jdk(r|St t |j|z|z }|dkr$|d|jtjdStj dd|jdtj }tj dd|dtj }tj|||jtjS) zLinear resample int16 PCM to ``to_rate``. Returns float32 samples in [-1, 1]. If the rates already match, this is a plain dtype cast. rrrNF)copyrq)numendpointr) rr6rsizerroundlinspacefloat64interp)rCrDrErn_outsrc_xdst_xs r=_resample_pcm16_to_raterSsll2::&0GGw||q0 gllW,y89 :E zv%%bjju%== KKSgllU"** UE KKSee2:: NE 99UE7 + 2 22:: >>rIrequested_modelc|xsdj}|r|Stjdxsdj}|xsdS)uPReturn the requested model ID, or the server-configured default. The default is read from ``$MLX_AUDIO_REALTIME_MODEL`` (settable via the ``--realtime-model`` CLI flag). Returns ``None`` when neither the client nor the server supplies a model — the caller is expected to reject the connection with a clear error. rMLX_AUDIO_REALTIME_MODELN)rrr)rT normalized configureds r=_resolve_realtime_model_namerYsH"'R..0J))67=2DDFJ  rIzmlx-community/silero-vad_REALTIME_VAD_MODEL_DEFAULT_realtime_vad_modelsc`tjdxsdjxstS)zReturn the VAD model used for server-side turn detection. Read at request time from ``$MLX_AUDIO_VAD_MODEL`` (settable via the ``--vad-model`` CLI flag), defaulting to Silero. MLX_AUDIO_VAD_MODELr)rrrrZrWrIr=_resolve_vad_model_namer^s* '(.B eg -,-rIr)cbtj|}|ddlm}||}|t|<|S)zLoad and cache a streaming VAD model for server-side turn detection. Kept behind a module-level indirection so the realtime endpoint can be exercised with a fake VAD model in tests without hitting Hugging Face. r)load)r[r mlx_audio.vadr`)r)rmload_vads r=_load_realtime_vad_modelrcs5 ! $ $T *E }2%*T" LrIz /v1/realtimec  !"#$%&'Kjd{dtfddtffd "dtf"fd }jj d}t | )|d d{j d{y tj td s-|d d d{j d{yd%t&t%&#gdtd'd}dtfd}d7dtdtf"#fd "fd!d8!"#%&f d }dt!j"j$dd$dtf $'fd }"d|dd{ j'd{} t)j*|} | j dd} | dk(r| j dxsi} | j d xsij d!xsi} | j d"xsi} | j dxs| j d}| j d#xsi}|j d$}t/|tr|d%kDr|t |}|rp| k7rk tj|td s|d |d d{D| t04d{t%&#dddd{gdd&| vr@ t3| j d&''d}n t7t9}t;|'}"d(|dd{n| d)k(r| j d d}|s%|#|"d*d+dd,dd-igd.d/d{t=j>tAjB|t<jD0}tG|#jH}t04d{#jK|dddd{d1d{|1tG|tL}t04d{|jO|}dddd{D]}|jPtRjTurD#|"d*d+dd,dd-igd.d/d{"d2|jVd3d{c"d4|jVd5d{|d{|jYnK| d6k(rF#|"d*d+dd,dd-igd.d/d{|d{||jY7O77#t$r7}|d |d{7j d{7Yd}~yd}~wwxYw77767#t(j,$r|dd{7Y]wxYw#t$r}|d |d{7Yd}~d}~wwxYw7O7?7$#1d{7swY5xYw#t4$r%}|t|d{7Yd}~d}~wwxYw#t$r#}d'd}|d'|d{7Yd}~d}~wwxYw7=777#1d{7swYxYw77j7L#1d{7swY]xYw777777#tZ$rYnTt$rI}d%dl.}|j_ |t|d{7n#t$rYnwxYwYd}~nd}~wwxYw j d{7n#t$rYnwxYwtajby# j d{7n#t$rYnwxYwtajbwxYww)9uUOpenAI Realtime API-compatible WebSocket endpoint. Works with any STT model whose ``load_model`` result exposes the streaming-session protocol. A streaming session must provide: - ``feed(samples: np.ndarray[float32])`` - ``close()`` - ``step(max_decode_tokens: int) -> list[str]`` - ``done: bool`` - ``input_sample_rate: int`` (the native rate expected by ``feed``) Models that support an adjustable transcription-delay / latency knob should accept ``transcription_delay_ms: Optional[int]`` on ``create_streaming_session``; models without the concept can ignore it. Client → server events follow OpenAI Realtime (subset): ``session.update``, ``input_audio_buffer.append``, ``input_audio_buffer.commit``. Server → client events: ``session.created`` / ``session.updated``, ``conversation.item.added``, ``input_audio_buffer.committed``, ``conversation.item.input_audio_transcription.delta`` / ``.completed``, and ``error``. Turn detection follows OpenAI's ``turn_detection`` on ``session.audio.input``. With ``{"type": "server_vad"}`` the server runs a streaming VAD (Silero by default, see ``--vad-model``), emits ``input_audio_buffer.speech_started`` / ``input_audio_buffer.speech_stopped`` and auto-commits each turn — the client never sends ``input_audio_buffer.commit``. With ``null`` (the default) the client drives commits manually. ``semantic_vad`` is not implemented yet. Model is selected via ``?model=`` or ``session.update.model``; defaults to ``$MLX_AUDIO_REALTIME_MODEL``. Nr-cJdtjjddS)Nevent_ruuiduuid4hexrWrIr= _new_event_idz"realtime_ws.._new_event_id1s# (("-.//rIrcZKdi|}j|d{y7w)Nevent_id)r)rrkrs r= send_eventzrealtime_ws..send_event4s,}:':!!'***s +)+rc<Kdd|idd{y7w)Nrr)r0rrW)rrns r= send_errorzrealtime_ws..send_error8s'Y4HIJJJs rmzsno realtime model configured: pass ?model=, set session.update.model, or start the server with --realtime-modelz load failed: r@zmodel z does not support streamingrHrzr=cJdtjjddS)Nitem_rrgrWrIr= _new_item_idz!realtime_ws.._new_item_id[s#tzz|'',-..rIrmax_decode_tokenscKt4d{j|}dddd{D](}j|dd|dd{*jS7_7?#1d{7swYOxYw7+w)uXRun one session.step, ship deltas. Returns session.done. MLX runs inline on the event-loop thread (not via a worker thread): MLX streams are thread-bound, so all realtime MLX work — the transcription step and the VAD — must share one thread, otherwise you hit "no Stream(gpu, N) in current thread". Nruz1conversation.item.input_audio_transcription.deltar)r0item_id content_indexr)REALTIME_INFERENCE_LOCKrrdr)rudeltasrcurrent_item_idfull_text_partsrnsessions r= drain_deltasz!realtime_ws..drain_deltas^s+*\\4E\FF+*E  " "5 )O.%&"   ||+**** sM B A.B A2 B A0*B BB 0B 2B8A; 9BB c^Kdj}dd|dd{y7w)Nrz5conversation.item.input_audio_transcription.completedr)r0rxry transcript)join)rr|r}rns r= send_donezrealtime_ws..send_donets8ww'O*!""     s "-+-c Kdddd{t4d{jdddd{dd{s dd{sd{t4d{tdddd{gdy777m#1d{7swY}xYw7s7c7U7H7,#1d{7swY.finalize_turns6*$(    +* MMO+*$r::: %r:::k**-;9OG+*# +****::*****sC&B,C&B.C&B2 C&B0C&CC&%C &C&+ C&5C 6C&C C&C C&"C# C&.C&0C&2C8B; 9CC& C& C& C&C&C#C C#C&sess_rc NddddddirjnddidS) Nzrealtime.session transcriptionrnz audio/pcm)r0raterm)rVrturn_detection)rrr0rmr)to_dict)client_input_raterJ session_id turn_configsr=_session_snapshotz&realtime_ws.._session_snapshotsH(#'2J N   oo))*5 54 56*/JKLLLooK<>% ;1GG"$O%)O5-1K,0M/#/c$,  6))#2./0J  t    /.&00I2A6E4=4:9?8O7P 1* %& '" !" !"#-,O6@6I6I/>!"##-,O4>4G4G/>!"##0/11)446A'2D88"*&2nO$$=&5*9(1(.-3],C+D %    $o%% ,!,,.}   =,---oo M|R1''  000 *%!(=)<=== ! 76666.!(Q000 ! )%*.K,0M",/@-D"EEE$ %  322228 76666 !"2 &    SV$ $ $     //# # #      //# # #    s_U4A_U7_U:_U=4_W_'W(B(_W_[4*W +[40W C[4#W78[4X"[4&X%'[4*X+9 [4X( [4Y,[42Y2[4$Z!%A[4)Z$*A[4Z'[4 Z- [4(Z*)[49[:"[4[[4 [ 2 [4=[ >A[4["[4![%"[4[([4[+<[4 [. [4[1[44_7_:_= V=V8VV8-V0.V83_8V==___ [4 W4*W-+W40[43W44[47 XXXX[4X[4%[4([4+X>1X4 2X>9[4 Y/ Y*Y!Y*$[4*Y//[42 Z;Z ZZ[4Z[4$[4'[4*[4-[3Z6 4[; [4[4 [4 [[ [ [4%[4([4+[4.[41[44 ]=^?]] \70\31\76] 7 ]] ]] ^ ]^].']*(].-_. ]:7_9]::__^.'^*(^.-_. ^:7_9^::__c"eZdZddZdZddZy)MLXAudioStudioServercx||_d|_t||_|jj dy)NT)exist_ok)start_ui ui_processrlog_dirmkdir)rGrrs r=rHzMLXAudioStudioServer.__init__as0  G}  D)rIclttjdz } tjddgt |tj tjd t|jd z d }tjgd t ||tj |_t!d |jdy#t$r tdtj$r}td|d}~wwxYw#t$r tdt$r}td|d}~wwxYw)z%Start UI with logs redirected to fileuinpminstallT)cwdstdoutstderrcheckuJ✗ Error: 'npm' is not installed or not found in PATH. UI will not start.u!✗ Error running 'npm install': Nui.logw)rrundev)rrru✓ UI started (logs: z/ui.log)uN✗ Error: 'npm' is not installed or not found in PATH. UI server not started.u✗ Failed to start UI: )r__file__parent subprocessrrUDEVNULLPIPEFileNotFoundErrorrCalledProcessErroropenrPopenSTDOUTrr)rGui_pathrui_logs r=start_ui_backgroundz(MLXAudioStudioServer.start_ui_backgroundgs$x.''$. E NN "L!))!   <$,,137F(..%L!(( DO *4<<.A B!! \ ,, E@!D D E! `  <6qc:; ; ? ~~,0NN () *589P9P5Q 12sD(() !4==$,, OF  YY YY{{ rI__main__rA)T)__doc__rrCrrrWrrrrrhr contextlibr dataclassesrrrpathlibrtypingrr r r urllib.parser mlx.corecorerbnumpyr6rrfastapir rrrrrrrrfastapi.middleware.corsrfastapi.responsesrrrhuggingface_hub.errorsrpydanticrmlx_audio.audio_iorrrrmlx_audio.realtime_vadrrr r!r"r#mlx_audio.server_inferencer$r%r&r'r(mlx_audio.tts.continuousr)r*mlx_audio.utilsr+r3r?rXrUrdrallowed_origins_envsplitrdefault_originsrjrouterlifespan_contextrlrrrrDrzrhrrrrrrrrrrrzrrrrrrrpostrdeleterNrrrVrr5rrrrrr9rr<rBrSrYrZr[r^rcrrrrR)origins0r=rs   *77,,    3PP:13C&334,,.i G d3i * bii ;<#6";";C"@A"@V\\^"@A  3 $G$$+ I, 29 2&',,..2(?+2      1#1CD2'(AB/3/dc5c5LJ 3J Z @!:@Fo<< K-1$Zo o o osm o  o  ooo oc]oo 3-ooo&c]o&od !"Cy:;H~z"b )1 )1 )1)1  )1  )1#)1fC$C$C$ C$ sm C$  C$C$L23dd4dV!&# 45 4HSM 4? ::?"%?03?ZZ?& (3- HSM $>S=')d38n)-- 3 ~OOOd IAIAXaH zFE<BsR