adityabanerjee/voxcpm2-studio
0
1o
2 ��wj� � @ s d dl Z e j�dd� e j�dd� d dlZd dlZd dlZd dlZ d dl3Z4d dl Z d dl m
Z
dZdZd Ze
jed5d6d d �Zefd
ededee fdd�Z dRd
edededB dedededefdd �Zejed � dSd
edededB dedededefd d ��Zd!Zej ej j d"d#�ed$���Z e� d%� e�!� �� ej"d&d'��) ej#d(d)d*d+d,�Z$ej#d-d d.�Z%ej&d/d0d1d2gd3�Z'ej(d4d5d6�Z)W d � n1 s�w Y ej"d7d'��C ej&d8d0d9�Z*ej+d:d7d;��$ ej,d<d=dd>d?d@�Z-ej,dAdBddCdDd@�Z.ej/dd dEdF�Z0W d � n 1 �sw Y W d � n 1 �s'w Y W d � n 1 �s7w Y e)j1ee$e%e'e-e.e0ge*dGdCdH� ej2dIgdJgdKgge$ge*edLdMdN� W d � n 1 �siw Y e3dOk�re j4dCdP�j5dLdQ� dS dS )T� NZPYTORCH_CUDA_ALLOC_CONFzexpandable_segments:TrueZTOKENIZERS_PARALLELISMZfalse)�VoxCPMzopenbmb/VoxCPM2i� i FZcuda)Z
load_denoiser�optimizeZdevice�text� max_chars�returnc C s� ddl }|�d| �� �}g }d}|D ](}|�� }|sq|r1t|�t|� d |kr1|�|� |}q|� d|� ��� }q|rB|�|� |S )zASplit long narration at sentence boundaries for stable synthesis.r Nu (?<=[.!?。！？])\s*� � � )�re�split�strip�len�append)r r r8 Z sentences�chunksZcurrentZsentence� r �hf_space_voxcpm2/app.py�9split_text s 1011r r � @�12 �* �voice_description�reference_audio� cfg_value�inference_timesteps�seedc O s0 t dtdtdt| p13d�d t|�d ���S )zBEstimate ZeroGPU reservation from text length and diffusion steps.�� �- �# r g� ��Q��?� )�min�max�intr
)r r r r r r �args�kwargsr r r �estimate_gpu_seconds1 s 0 r$ )Zduration�,A warm, confident narrator with clear pacingc C s� d� | pd�� �}|st�d��t|�tkr t�dt� d���t|�}t|�}t|�}d| kr5dks;n t�d��d | krEd14ksKn t�d ��t|�}g }t |�D ]2\} }15|16} |sk|�17� rkd |�18� � d
|19� �} t j | |pqd||ddd|| d�} |�
tj| tjd�� qU|s�t�d��tjtt jjd �tjd�}
g }t |�D ]\} } | r�|�
|
� |�
| � q�t�|�}tjddd�}|�� tj|j|t jjdd� |jS )a� Generate 48 kHz speech with VoxCPM2.20 21 Args:22 text: Narration to synthesize in any VoxCPM2-supported language.23 voice_description: Natural-language voice style used without reference audio.24 reference_audio: Optional short WAV recording for voice cloning.25 cfg_value: Guidance strength, normally between 1.5 and 3.0.26 inference_timesteps: Diffusion steps; higher can improve quality but costs time.27 seed: Reproducibility seed.28 r r zEnter text to synthesize.zText is limited to z characters per request.� �?� @z CFG must be between 1.0 and 4.0.� � z)Inference steps must be between 4 and 20.�(�)NTF)r Zreference_wav_pathr r � normalizeZdenoiseZ
retry_badcaser )Zdtypez VoxCPM2 did not produce audio.g29ףp=30�?z.wav)�delete�suffixZPCM_16)Zsubtype)�joinr �gr�Errorr
�MAX_TEXT_CHARS�floatr! r � enumerater �model�generater �npZasarrayZfloat32ZzerosZ tts_modelZ sample_rateZ concatenate�tempfileZNamedTemporaryFile�close�sf�write�name)r r r r r r Zcleanedr Z generated�index�chunkZ
designed_textZwav�pause�piecesZfull_wav�outputr r r �31synthesize? sT 32 3334 �3536 37 38rB z�39.gradio-container { max-width: 1040px !important; margin: auto !important; }40.hero { text-align: center; padding: 1rem 0 .25rem; }41.hero h1 { font-size: 2.3rem; margin-bottom: .35rem; }42.dark .gradio-container { color: var(--body-text-color); }43Zindigo)Z primary_hue)ZthemeZcssu� <div class='hero'><h1>🎙️ VoxCPM2 Studio</h1><p>Multilingual 48 kHz speech, voice design, and optional voice cloning.</p></div>r )ZscaleZ Narration� � uK Type narration in English, Hindi, Chinese, or another supported language…)�label�linesZ max_linesZ placeholderz Voice design)rE �valuezReference voice (optional)�filepathZuploadZ44microphone)rE �typeZsourceszGenerate speechZprimary)Zvariant� zGenerated 48 kHz speech)rE rI zAdvanced settings)�openr& r' g�������?z CFG guidance)rG �steprE r( r) r zInference stepsZSeed)rG Z precisionrE Ztts)�fn�inputs�outputsZapi_nameZconcurrency_limitz[VoxCPM2 brings multilingual speech and expressive voice design to open-source applications.u� नमस्ते! यह एक स्वाभाविक और स्पष्ट हिंदी आवाज़ का उदाहरण है।u- 你好，欢迎体验多语言语音合成。TZlazy)ZexamplesrN rO rM Zcache_examplesZ45cache_mode�__main__)Zdefault_concurrency_limit)Z46mcp_server)r Nr r r )r% Nr r r )6�os�environ�47setdefault�spacesZgradior0 Znumpyr7 Z soundfiler: r8 ZtorchZvoxcpmr ZMODEL_IDr2 ZMAX_CHUNK_CHARSZfrom_pretrainedr5 �strr! �listr r3 r$ ZGPUrB ZCSSZBlocksZthemesZSoftZdemoZHTMLZRowZColumnZTextboxr r ZAudior ZButtonr6 Z output_audioZ AccordionZSliderZcfgZsteps�Numberr ZclickZExamples�__name__�queueZlaunchr r r r �<module> s� � ������� 48�49������� �B�50��51� � �� �� ���� �521�