o 6h/ @sddlZddlmmZddlmZmZddlZ ddl Z ddl Z ddl mZddlZddlmZmZmZddlZddlmZddlZddlZddlZGdddZdadefd d Zdd ed edeeeffddZdS)N) VitsModel AutoTokenizer)Path)OptionalUnionList)ThreadPoolExecutorc @seZdZdZ     d.dedeeded ed ef d d Zdeed e j fddZ ddZ ddZ ddZded efddZdedejfddZ  d/dededed ejfdd Zd!ejd ejfd"d#Zd0ded%ed efd&d'Zd0ded%ed efd(d)Zd*d+Zd efd,d-ZdS)1HindiTTSProcessorzz Professional Hindi Text-to-Speech Processing System Optimized for speed, accuracy, and production deployment facebook/mms-tts-hinNTd model_namedeviceenable_cuda_optimization cache_sizeuse_half_precisioncCsh||_||_i|_tj|d|_|||_|o|jjdk|_ | |r.t j r.||dS)a Initialize the TTS system with performance optimizations Args: model_name: HuggingFace model identifier device: Target device ('cuda', 'cpu', or None for auto-detection) enable_cuda_optimization: Enable CUDA-specific optimizations cache_size: Size of the audio cache for repeated texts use_half_precision: Use FP16 for faster inference (if supported) )maxsizecudaN)r r audio_cachequeueQueue cache_queue _setup_devicer typer _load_modelstorchr is_available_apply_cuda_optimizations _warmup_model)selfr r rrrrL/var/www/eduai.edurigo.com/doc_train/edurigo_ai/text_to_speech/processing.py__init__s  zHindiTTSProcessor.__init__returncCsR|durtjr d}nd}t|}|jdkr'tjdtjj_dtjj_ |S)z"Setup optimal device configurationNrcpuTF) rrrr r empty_cachebackendscudnn benchmark deterministic)rr device_objrrr r:s      zHindiTTSProcessor._setup_devicecCsptj|jddd|_tj|j|jrtjntjddd|_ |j |j |_ |j t |j dr6d|j j_dSdS)z#Load models with optimization flagsTz.cache/tokenizers)use_fast cache_dirz .cache/models) torch_dtypelow_cpu_mem_usager+configN)rfrom_pretrainedr tokenizerrrrfloat16float32modeltor evalhasattrr. use_cacherrrr rLs   zHindiTTSProcessor._load_modelsc CstjrAttdr(z tj|jdd|_Wnty'}zWYd}~nd}~wwt|jdrCz |jjtjd|_WdSYdSdSdS)z!Apply CUDA-specific optimizationscompilez max-autotune)modeNr4) memory_format) rrrr6r9r3 Exceptionr4 channels_last)rerrr res     z+HindiTTSProcessor._apply_cuda_optimizationscCsd}t'|j|dd|j}|jrdd|D}|jdi|}Wdn1s0wY|jjdkrBtj dSdS) z'Warmup model for consistent performanceuनमस्तेpt)return_tensorscS*i|]\}}||jtjkr|n|qSrdtyperr2half.0kvrrr }s*z3HindiTTSProcessor._warmup_model..Nrr) rno_gradr0r4r ritemsr3rrr$)r warmup_textinputs_rrr rvs  zHindiTTSProcessor._warmup_modeltextcCsdt|S)zGenerate cache key for texttts_)hashstriplower)rrOrrr _get_cache_keysz HindiTTSProcessor._get_cache_keykey audio_datacCszt|j|jkr#z|j}||jvr|j|=Wn tjy"Ynw||j|<z |j|WdStjy<YdSw)z)Manage audio cache with LRU-like behaviorN) lenrrr get_nowaitrEmpty put_nowaitFull)rrUrV oldest_keyrrr _manage_caches   zHindiTTSProcessor._manage_cacher7normalize_audioc Cs|r|s td|}|r||nd}|r#||jvr#|j|St|j|ddddd|j}|j rBdd| D}|j rktj rktj jd dd |jdi|}Wdn1sewYn|jdi|}t|d r||j}n(t|d r|j}nt|D]}t||} t| tjr| jd kr| }nqtd|} |r|| } Wdn1swY|r|r||| | S)aD Synthesize speech from Hindi text with maximum performance Args: text: Hindi text to synthesize use_cache: Use audio cache for repeated texts normalize_audio: Normalize output audio Returns: Audio waveform as numpy array zText cannot be emptyNr?Ti)r@padding truncation max_lengthcSrArrBrErrr rIsz0HindiTTSProcessor.synthesize..r)enabledwaveformaudioz-Could not find audio output in model responser)rR ValueErrorrTrrrJr0r4r rrKrrampautocastr3r6rcrddirgetattr isinstanceTensorndim RuntimeErrorsqueezer#floatnumpy_normalize_audior]) rrOr7r^ cache_keyrMoutputrc attr_name attr_valueaudio_nprrr synthesizesZ         , zHindiTTSProcessor.synthesizerdcCs6|t|}tt|}|dkr|d|}|S)z@Normalize audio to prevent clipping and ensure consistent volumerg?)npmeanmaxabs)rrdmax_valrrr rrs  z"HindiTTSProcessor._normalize_audio> sample_ratec Csz(||}|jtjkr|tj}t}tj|||dd| d| WSt yD}zt dt |jdt|d}~ww)z Synthesize speech and return as bytes for API response Args: text: Hindi text to synthesize sample_rate: Audio sample rate Returns: Audio data as bytes WAV)formatrzError in synthesize_to_bytes: z: N)rxrCryr2astypeioBytesIOsfwriteseekgetvaluer<printr__name__str)rrOrrwbufferr>rrr synthesize_to_bytess     z%HindiTTSProcessor.synthesize_to_bytescCs|||}t|dS)z Synthesize speech and return as base64 encoded string Args: text: Hindi text to synthesize sample_rate: Audio sample rate Returns: Base64 encoded audio data zutf-8)rbase64 b64encodedecode)rrOr audio_bytesrrr synthesize_to_base64s z&HindiTTSProcessor.synthesize_to_base64cCsL|j|js$z|jWn tjyYdSw|jr dSdS)z Clear audio cache to free memoryN)rclearremptyrXrrYr8rrr clear_caches  zHindiTTSProcessor.clear_cachec CsZt|j|jt|j|j|jtj tj rtj ndtj r)tj dSddS)zGet performance statisticsr)r half_precisionrmax_cache_sizer cuda_availablecuda_memory_allocatedcuda_memory_reserved) rr rrWrrr rrrmemory_allocatedmemory_reservedr8rrr get_stats%szHindiTTSProcessor.get_stats)r NTr T)TT)r~)r __module__ __qualname____doc__rrboolintr!rr rrrrrTryndarrayr]rxrrbytesrrrdictrrrrr r sP # K  r r"cCstdur tddddatS)z/Get or create the global TTS processor instanceNT2)rrr)_tts_processorr rrrr get_tts_processor6srrrO output_formatcCs"t}|dkr ||S||S)z Main processing function for text to speech conversion Args: text: Hindi text to convert to speech output_format: Output format - 'bytes' or 'base64' Returns: Audio data in requested format r)rrr)rOr processorrrr process_text_to_speechAs   r)r) rtorch.nn.functionalnn functionalF transformersrr soundfilerrqrytimepathlibrgctypingrrr threadingconcurrent.futuresrrrrr rrrrrrrrr s(  %$