whisperer-local.py

import sounddevice as sd
import numpy as np
import openai
import pyperclip
from pynput.keyboard import Listener, Controller, Key, KeyCode
from scipy.io.wavfile import write
import whisper

# Path to the file containing your OpenAI API key
api_key_path = 'openai_api_key.txt'

# Key to hold down to start recording
# record_key = Key.alt_r
record_key = Key.ctrl_r

# Key to tap turn on translation
# translate_key = KeyCode.from_char('t')
translate_key = Key.shift_r

# Initialize variables
recording = False
audio_data = []
stream = None
translate = False
# If true, the transcript will be copied to the clipboard even it doesn't contain any special characters.
force_clipboard = False

model = whisper.load_model("small.en")

# Print a nice message if the API key file isn't present.
try:
    with open(api_key_path, 'r') as file:
        pass
except FileNotFoundError:
    print(f"Please create a file called {api_key_path} and paste your OpenAI API key in it.")
    exit()

# Read the API key from the file
with open(api_key_path, 'r') as file:
    openai.api_key = file.read().strip()

# Callback function to collect audio data
def callback(indata, frames, time, status):
    global audio_data
    if recording:
        # Check if indata has the expected shape (e.g., (32,))
        if indata.shape[1] == 1:  # Check if indata is mono
          audio_data.append(indata.copy())
        else:
          # Resize indata or discard it
          pass

keyboard = Controller()

def on_press(key):
    global recording, stream, audio_data, translate

    if key == record_key and not recording:
      recording = True
      translate = False
      print("Recording started...")

      audio_data = []
      
      # Initialize and start InputStream
      stream = sd.InputStream(callback=callback, channels=1, samplerate=16000)
      stream.start()

    # If recording and the translate key is pressed, set translate to True and throw away the keypress.
    if recording and key == translate_key:
      print("Translate key pressed.")
      translate = True
      
def on_release(key):
    global recording, stream, translate, force_clipboard

    if key == record_key:
      recording = False
      print("Recording stopped.")

      # Stop and close InputStream
      if stream is not None:
          stream.stop()
          stream.close()
          stream = None

      if audio_data == []:
        print("No audio data recorded.")
        return
      
      # Concatenate all audio data into one NumPy array
      audio_data_np = np.concatenate(audio_data, axis=0)

      # Get length of audio data in seconds
      audio_data_length = len(audio_data_np) / 16000

      # If was start of a double-click, set force_clipboard to True
      if audio_data_length < 0.5:
        force_clipboard = True

      if audio_data_length < 1:
        print("Audio data is less than 1 second long.")
        return

      # Write audio data to file
      write("output.wav", 16000, audio_data_np)
      
      result = model.transcribe("output.wav", initial_prompt="How are you doing today? I'm really looking forward to seeing you again!", fp16=False)
      transcript_text = result["text"]

      # Replace "New paragraph." with "\n"
      transcript_text = transcript_text.replace("New paragraph.", "\n\n")

      # Remove initial whitespace
      transcript_text = transcript_text.strip()

      print("Transcript:")
      print(transcript_text)

      if translate:
        translate = False

        print("Translating transcript to French...")
        result = openai.ChatCompletion.create(
          model="gpt-4",
          messages=[
            {"role": "system", "content": "You translate the input text to Quebec French. You only output the text and nothing else."},
            # {"role": "system", "content": "You take the input and make it polite, business appropriate, and kind. You only output the text and nothing else."},
            {"role": "user", "content": transcript_text},
          ]
        )

        transcript_text = result["choices"][0]["message"]["content"]
        print(transcript_text)

      # Determine if any special characters are being used that can't be
      # typed using keyboard.type(). These are any characters that aren't in English
      allowed_chars = set('abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789 .,;:!?-\'_')
      special_chars = set(transcript_text) - allowed_chars
      if len(special_chars) > 0 or force_clipboard:
        print("Special characters detected: " + str(special_chars))

        # Copy the transcript text to the clipboard
        pyperclip.copy(transcript_text)

        # Simulate CTRL-V to paste the text
        keyboard.press(Key.ctrl)
        keyboard.press('v')
        keyboard.release('v')
        keyboard.release(Key.ctrl)

        force_clipboard = False
      else:  
        # Since there are no accents, we can just use the standard type command.
        keyboard.type(transcript_text)
      
# Start listening for key events
with Listener(on_press=on_press, on_release=on_release) as listener:
    listener.join()