o i@sddlZddlZddlZddlmZddlZddlmZm Z ddl m Z ddl m Z ddlmZddlmZedd Zed d Zd ed efddZeddefddZdS)N) Generator)FastAPI HTTPException)Response)write)Tensor) SopranoTTSd) cache_size_mbzSoprano TTS API)titletensorreturncCsDt|dddtj}t}t|d||d| S)z> Convert a 1D fp32 torch tensor to a WAV byte stream. gg?ii}r) npclipnumpyastypeint16ioBytesIOrseekread)r audio_int16wav_iorB/home/ubuntu/.local/lib/python3.10/site-packages/soprano/server.py_tensor_to_wav_bytess   rz/v1/audio/speechpayloadcsN|d}t|tr|stdddt|}t|}t|dddidS) z Minimal implementation of OpenAI's Speech endpoint. Fields: - input: string - text to synthesize - model, voice, etc. are accepted but ignored. - response_format: str - ignored, only support wav. inputiz)`input` field must be a non-empty string.) status_codedetailz audio/wavzContent-Dispositionz!attachment; filename="speech.wav")content media_typeheaders) get isinstancestrstriprttsinferrr)rtext audio_tensor wav_bytesrrr create_speech s   r,)base64rjsontypingrrrfastapirrfastapi.responsesrscipy.io.wavfilertorchr soprano.ttsrr'appbytesrpostdictr,rrrrs