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.
252 lines
8.9 KiB
252 lines
8.9 KiB
# This is a websocket interpreter, TTS and STT disabled.
|
|
# It makes a websocket on port 8000 that sends/recieves LMC messages in *streaming* format.
|
|
|
|
### You MUST send a start and end flag with each message! For example: ###
|
|
|
|
"""
|
|
{"role": "user", "type": "message", "start": True})
|
|
{"role": "user", "type": "message", "content": "hi"})
|
|
{"role": "user", "type": "message", "end": True})
|
|
"""
|
|
|
|
###
|
|
from pynput import keyboard
|
|
from .utils.bytes_to_wav import bytes_to_wav
|
|
from RealtimeTTS import TextToAudioStream, CoquiEngine, OpenAIEngine, ElevenlabsEngine
|
|
from RealtimeSTT import AudioToTextRecorder
|
|
import time
|
|
import asyncio
|
|
import json
|
|
import os
|
|
|
|
|
|
class AsyncInterpreter:
|
|
def __init__(self, interpreter, debug):
|
|
self.stt_latency = None
|
|
self.tts_latency = None
|
|
self.interpreter_latency = None
|
|
# time from first put to first yield
|
|
self.tffytfp = None
|
|
self.debug = debug
|
|
|
|
self.interpreter = interpreter
|
|
self.audio_chunks = []
|
|
|
|
# STT
|
|
self.stt = AudioToTextRecorder(
|
|
model="tiny.en", spinner=False, use_microphone=False
|
|
)
|
|
|
|
self.stt.stop() # It needs this for some reason
|
|
|
|
# TTS
|
|
if self.interpreter.tts == "coqui":
|
|
engine = CoquiEngine()
|
|
elif self.interpreter.tts == "openai":
|
|
engine = OpenAIEngine()
|
|
elif self.interpreter.tts == "elevenlabs":
|
|
engine = ElevenlabsEngine(api_key=os.environ["ELEVEN_LABS_API_KEY"])
|
|
engine.set_voice("Michael")
|
|
else:
|
|
raise ValueError(f"Unsupported TTS engine: {self.interpreter.tts}")
|
|
self.tts = TextToAudioStream(engine)
|
|
|
|
self.active_chat_messages = []
|
|
|
|
self._input_queue = asyncio.Queue() # Queue that .input will shove things into
|
|
self._output_queue = asyncio.Queue() # Queue to put output chunks into
|
|
self._last_lmc_start_flag = None # Unix time of last LMC start flag recieved
|
|
self._in_keyboard_write_block = (
|
|
False # Tracks whether interpreter is trying to use the keyboard
|
|
)
|
|
self.loop = asyncio.get_event_loop()
|
|
|
|
async def _add_to_queue(self, queue, item):
|
|
await queue.put(item)
|
|
|
|
async def clear_queue(self, queue):
|
|
while not queue.empty():
|
|
await queue.get()
|
|
|
|
async def clear_input_queue(self):
|
|
await self.clear_queue(self._input_queue)
|
|
|
|
async def clear_output_queue(self):
|
|
await self.clear_queue(self._output_queue)
|
|
|
|
async def input(self, chunk):
|
|
"""
|
|
Expects a chunk in streaming LMC format.
|
|
"""
|
|
if isinstance(chunk, bytes):
|
|
# It's probably a chunk of audio
|
|
self.stt.feed_audio(chunk)
|
|
self.audio_chunks.append(chunk)
|
|
# print("INTERPRETER FEEDING AUDIO")
|
|
|
|
else:
|
|
|
|
try:
|
|
chunk = json.loads(chunk)
|
|
except:
|
|
pass
|
|
|
|
if "start" in chunk:
|
|
# print("Starting STT")
|
|
self.stt.start()
|
|
self._last_lmc_start_flag = time.time()
|
|
# self.interpreter.computer.terminal.stop() # Stop any code execution... maybe we should make interpreter.stop()?
|
|
elif "end" in chunk:
|
|
# print("Running OI on input")
|
|
asyncio.create_task(self.run())
|
|
else:
|
|
await self._add_to_queue(self._input_queue, chunk)
|
|
|
|
def add_to_output_queue_sync(self, chunk):
|
|
"""
|
|
Synchronous function to add a chunk to the output queue.
|
|
"""
|
|
# print("ADDING TO QUEUE:", chunk)
|
|
asyncio.create_task(self._add_to_queue(self._output_queue, chunk))
|
|
|
|
def generate(self, message, start_interpreter):
|
|
last_lmc_start_flag = self._last_lmc_start_flag
|
|
self.interpreter.messages = self.active_chat_messages
|
|
|
|
# print("message is", message)
|
|
|
|
for chunk in self.interpreter.chat(message, display=True, stream=True):
|
|
|
|
if self._last_lmc_start_flag != last_lmc_start_flag:
|
|
# self.beeper.stop()
|
|
break
|
|
|
|
# self.add_to_output_queue_sync(chunk) # To send text, not just audio
|
|
|
|
content = chunk.get("content")
|
|
|
|
# Handle message blocks
|
|
if chunk.get("type") == "message":
|
|
if content:
|
|
# self.beeper.stop()
|
|
|
|
# Experimental: The AI voice sounds better with replacements like these, but it should happen at the TTS layer
|
|
# content = content.replace(". ", ". ... ").replace(", ", ", ... ").replace("!", "! ... ").replace("?", "? ... ")
|
|
# print("yielding ", content)
|
|
if self.tffytfp is None:
|
|
self.tffytfp = time.time()
|
|
|
|
yield content
|
|
|
|
# Handle code blocks
|
|
elif chunk.get("type") == "code":
|
|
if "start" in chunk:
|
|
# self.beeper.start()
|
|
pass
|
|
|
|
# Experimental: If the AI wants to type, we should type immediatly
|
|
if (
|
|
self.interpreter.messages[-1]
|
|
.get("content", "")
|
|
.startswith("computer.keyboard.write(")
|
|
):
|
|
keyboard.controller.type(content)
|
|
self._in_keyboard_write_block = True
|
|
if "end" in chunk and self._in_keyboard_write_block:
|
|
self._in_keyboard_write_block = False
|
|
# (This will make it so it doesn't type twice when the block executes)
|
|
if self.interpreter.messages[-1]["content"].startswith(
|
|
"computer.keyboard.write("
|
|
):
|
|
self.interpreter.messages[-1]["content"] = (
|
|
"dummy_variable = ("
|
|
+ self.interpreter.messages[-1]["content"][
|
|
len("computer.keyboard.write(") :
|
|
]
|
|
)
|
|
|
|
# Send a completion signal
|
|
if self.debug:
|
|
end_interpreter = time.time()
|
|
self.interpreter_latency = end_interpreter - start_interpreter
|
|
print("INTERPRETER LATENCY", self.interpreter_latency)
|
|
# self.add_to_output_queue_sync({"role": "server","type": "completion", "content": "DONE"})
|
|
|
|
async def run(self):
|
|
"""
|
|
Runs OI on the audio bytes submitted to the input. Will add streaming LMC chunks to the _output_queue.
|
|
"""
|
|
self.interpreter.messages = self.active_chat_messages
|
|
|
|
self.stt.stop()
|
|
|
|
input_queue = []
|
|
while not self._input_queue.empty():
|
|
input_queue.append(self._input_queue.get())
|
|
|
|
if self.debug:
|
|
start_stt = time.time()
|
|
message = self.stt.text()
|
|
end_stt = time.time()
|
|
self.stt_latency = end_stt - start_stt
|
|
print("STT LATENCY", self.stt_latency)
|
|
|
|
if self.audio_chunks:
|
|
audio_bytes = bytearray(b"".join(self.audio_chunks))
|
|
wav_file_path = bytes_to_wav(audio_bytes, "audio/raw")
|
|
print("wav_file_path ", wav_file_path)
|
|
self.audio_chunks = []
|
|
else:
|
|
message = self.stt.text()
|
|
|
|
print(message)
|
|
|
|
# Feed generate to RealtimeTTS
|
|
self.add_to_output_queue_sync(
|
|
{"role": "assistant", "type": "audio", "format": "bytes.wav", "start": True}
|
|
)
|
|
start_interpreter = time.time()
|
|
text_iterator = self.generate(message, start_interpreter)
|
|
|
|
self.tts.feed(text_iterator)
|
|
if not self.tts.is_playing():
|
|
self.tts.play_async(on_audio_chunk=self.on_tts_chunk, muted=True)
|
|
|
|
while True:
|
|
await asyncio.sleep(0.1)
|
|
# print("is_playing", self.tts.is_playing())
|
|
if not self.tts.is_playing():
|
|
self.add_to_output_queue_sync(
|
|
{
|
|
"role": "assistant",
|
|
"type": "audio",
|
|
"format": "bytes.wav",
|
|
"end": True,
|
|
}
|
|
)
|
|
if self.debug:
|
|
end_tts = time.time()
|
|
self.tts_latency = end_tts - self.tts.stream_start_time
|
|
print("TTS LATENCY", self.tts_latency)
|
|
self.tts.stop()
|
|
|
|
break
|
|
|
|
async def _on_tts_chunk_async(self, chunk):
|
|
# print("adding chunk to queue")
|
|
if self.debug and self.tffytfp is not None and self.tffytfp != 0:
|
|
print(
|
|
"time from first yield to first put is ",
|
|
time.time() - self.tffytfp,
|
|
)
|
|
self.tffytfp = 0
|
|
await self._add_to_queue(self._output_queue, chunk)
|
|
|
|
def on_tts_chunk(self, chunk):
|
|
# print("ye")
|
|
asyncio.run(self._on_tts_chunk_async(chunk))
|
|
|
|
async def output(self):
|
|
# print("outputting chunks")
|
|
return await self._output_queue.get()
|