You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
01/software/source/server/services/stt/openai/stt.py

109 lines
3.3 KiB

class Stt:
def __init__(self, config):
pass
def stt(self, audio_file_path):
return stt(audio_file_path)
from datetime import datetime
import os
import contextlib
import tempfile
import ffmpeg
import subprocess
import openai
from openai import OpenAI
client = OpenAI()
def convert_mime_type_to_format(mime_type: str) -> str:
if mime_type == "audio/x-wav" or mime_type == "audio/wav":
return "wav"
if mime_type == "audio/webm":
return "webm"
if mime_type == "audio/raw":
return "dat"
return mime_type
@contextlib.contextmanager
def export_audio_to_wav_ffmpeg(audio: bytearray, mime_type: str) -> str:
temp_dir = tempfile.gettempdir()
# Create a temporary file with the appropriate extension
input_ext = convert_mime_type_to_format(mime_type)
input_path = os.path.join(temp_dir, f"input_{datetime.now().strftime('%Y%m%d%H%M%S%f')}.{input_ext}")
with open(input_path, 'wb') as f:
f.write(audio)
# Check if the input file exists
assert os.path.exists(input_path), f"Input file does not exist: {input_path}"
# Export to wav
output_path = os.path.join(temp_dir, f"output_{datetime.now().strftime('%Y%m%d%H%M%S%f')}.wav")
if mime_type == "audio/raw":
ffmpeg.input(
input_path,
f='s16le',
ar='16000',
ac=1,
).output(output_path, loglevel='panic').run()
else:
ffmpeg.input(input_path).output(output_path, acodec='pcm_s16le', ac=1, ar='16k', loglevel='panic').run()
try:
yield output_path
finally:
os.remove(input_path)
os.remove(output_path)
def run_command(command):
result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
return result.stdout, result.stderr
def get_transcription_file(wav_file_path: str):
local_path = os.path.join(os.path.dirname(__file__), 'local_service')
whisper_rust_path = os.path.join(os.path.dirname(__file__), 'whisper-rust', 'target', 'release')
model_name = os.getenv('WHISPER_MODEL_NAME', 'ggml-tiny.en.bin')
output, error = run_command([
os.path.join(whisper_rust_path, 'whisper-rust'),
'--model-path', os.path.join(local_path, model_name),
'--file-path', wav_file_path
])
return output
def get_transcription_bytes(audio_bytes: bytearray, mime_type):
with export_audio_to_wav_ffmpeg(audio_bytes, mime_type) as wav_file_path:
return get_transcription_file(wav_file_path)
def stt_bytes(audio_bytes: bytearray, mime_type="audio/wav"):
with export_audio_to_wav_ffmpeg(audio_bytes, mime_type) as wav_file_path:
return stt_wav(wav_file_path)
def stt_wav(wav_file_path: str):
audio_file = open(wav_file_path, "rb")
try:
transcript = client.audio.transcriptions.create(
model="whisper-1",
file=audio_file,
response_format="text"
)
except openai.BadRequestError as e:
print(f"openai.BadRequestError: {e}")
return None
return transcript
def stt(input_data, mime_type="audio/wav"):
if isinstance(input_data, str):
return stt_wav(input_data)
elif isinstance(input_data, bytearray):
return stt_bytes(input_data, mime_type)
else:
raise ValueError("Input data should be either a path to a wav file (str) or audio bytes (bytearray)")