| from transformers import AutoProcessor, SeamlessM4Tv2ForSpeechToText |
| import torchaudio |
| from torch import Tensor |
| from base64 import b64encode |
|
|
|
|
| class EndpointHandler(): |
| def __init__(self, path=""): |
| self.path = path |
| self.processor = AutoProcessor.from_pretrained( |
| "facebook/seamless-m4t-v2-large") |
| self.model = SeamlessM4Tv2ForSpeechToText.from_pretrained( |
| "facebook/seamless-m4t-v2-large") |
| self.AUDIO_DEFAULT_SAMPLE_RATE = 16000 |
|
|
| def __call__(self, data: dict): |
| input_audio_tensor = data["inputs"]["audio"] |
| input_lang_code = data["inputs"]["input_lang_code"] |
| output_lang_code = data["inputs"]["output_lang_code"] |
| audio_inputs = self.processor( |
| audios=input_audio_tensor, src_lang=input_lang_code, return_tensors="pt") |
| id_array = self.model.generate( |
| **audio_inputs, tgt_lang=output_lang_code) |
| text_array_from_audio = self.processor.batch_decode( |
| id_array, skip_special_tokens=True) |
| return [{"text_array": text_array_from_audio}] |
|
|
