SOTA Open Source TTS
from http import HTTPStatus
import numpy as np
from kui.asgi import HTTPException
from fish_speech.inference_engine import TTSInferenceEngine
from fish_speech.models.text2semantic.errors import InvalidInputError
from fish_speech.utils.schema import ServeTTSRequest
AMPLITUDE = 32768 # Needs an explaination
def inference_wrapper(req: ServeTTSRequest, engine: TTSInferenceEngine):
"""
Wrapper for the inference function.
Used in the API server.
"""
count = 0
for result in engine.inference(req):
match result.code:
case "header":
# inference_async passes on bytes only.
if isinstance(result.audio, tuple):
yield result.audio[1].tobytes()
case "error":
# Only input the model rejects is the client's fault; anything else,
# including other ValueErrors, is a server error.
raise HTTPException(
(
HTTPStatus.BAD_REQUEST
if isinstance(result.error, InvalidInputError)
else HTTPStatus.INTERNAL_SERVER_ERROR
),
content=str(result.error),
)
case "segment":
count += 1
if isinstance(result.audio, tuple):
# Clip so that 1.0 does not wrap around to -32768.
samples = np.clip(
result.audio[1] * AMPLITUDE, -AMPLITUDE, AMPLITUDE - 1
)
yield samples.astype(np.int16).tobytes()
case "final":
count += 1
if isinstance(result.audio, tuple):
yield result.audio[1]
return None # Stop the generator
if count == 0:
raise HTTPException(
HTTPStatus.INTERNAL_SERVER_ERROR,
content="No audio generated, please check the input text.",
)