RealTimeSTT 0.3.92__tar.gz → 0.3.94__tar.gz

{realtimestt-0.3.92 → realtimestt-0.3.94}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: RealtimeSTT
-Version: 0.3.92
+Version: 0.3.94
 Summary: A fast Voice Activity Detection and Transcription System
 Home-page: https://github.com/KoljaB/RealTimeSTT
 Author: Kolja Beigel
@@ -13,15 +13,15 @@ Requires-Python: >=3.6
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: PyAudio==0.2.14
-Requires-Dist: faster-whisper==1.1.0
+Requires-Dist: faster-whisper==1.1.1
 Requires-Dist: pvporcupine==1.9.5
 Requires-Dist: webrtcvad-wheels==2.0.14
 Requires-Dist: halo==0.0.31
 Requires-Dist: torch
 Requires-Dist: torchaudio
-Requires-Dist: scipy==1.14.1
+Requires-Dist: scipy==1.15.1
 Requires-Dist: openwakeword>=0.4.0
-Requires-Dist: websockets==14.1
+Requires-Dist: websockets==14.2
 Requires-Dist: websocket-client==1.8.0
 Requires-Dist: numpy<2.0.0
 

{realtimestt-0.3.92 → realtimestt-0.3.94}/RealtimeSTT/audio_input.py

@@ -1,22 +1,35 @@
 from colorama import init, Fore, Style
+from scipy.signal import butter, filtfilt, resample_poly
 import pyaudio
 import logging
-import time
 
+DESIRED_RATE = 16000
+CHUNK_SIZE = 1024
+AUDIO_FORMAT = pyaudio.paInt16
+CHANNELS = 1
 
 class AudioInput:
-    def __init__(self, input_device_index=None, debug_mode=False):
+    def __init__(
+            self,
+            input_device_index: int = None,
+            debug_mode: bool = False,
+            target_samplerate: int = DESIRED_RATE,
+            chunk_size: int = CHUNK_SIZE,
+            audio_format: int = AUDIO_FORMAT,
+            channels: int = CHANNELS,
+            resample_to_target: bool = True,
+        ):
+
         self.input_device_index = input_device_index
         self.debug_mode = debug_mode
         self.audio_interface = None
         self.stream = None
         self.device_sample_rate = None
-
-        # PyAudio constants
-        self.CHUNK = 1024
-        self.FORMAT = pyaudio.paInt16  
-        self.CHANNELS = 1
-        self.DESIRED_RATE = 16000
+        self.target_samplerate = target_samplerate
+        self.chunk_size = chunk_size
+        self.audio_format = audio_format
+        self.channels = channels
+        self.resample_to_target = resample_to_target
 
     def get_supported_sample_rates(self, device_index):
         """Test which standard sample rates are supported by the specified device."""
@@ -31,8 +44,8 @@ class AudioInput:
                 if self.audio_interface.is_format_supported(
                     rate,
                     input_device=device_index,  # Changed to input_device
-                    input_channels=max_channels,  # Changed to input_channels 
-                    input_format=self.FORMAT,  # Changed to input_format
+                    input_channels=max_channels,  # Changed to input_channels
+                    input_format=self.audio_format,  # Changed to input_format
                 ):
                     supported_rates.append(rate)
             except:
@@ -108,18 +121,18 @@ class AudioInput:
             if self.debug_mode:
                 print(f"Actual selected device index: {actual_device_index}")
             self.input_device_index = actual_device_index
-            self.device_sample_rate = self._get_best_sample_rate(actual_device_index, self.DESIRED_RATE)
+            self.device_sample_rate = self._get_best_sample_rate(actual_device_index, self.target_samplerate)
 
             if self.debug_mode:
                 print(f"Setting up audio on device {self.input_device_index} with sample rate {self.device_sample_rate}")
 
             try:
                 self.stream = self.audio_interface.open(
-                    format=self.FORMAT,
-                    channels=self.CHANNELS,
+                    format=self.audio_format,
+                    channels=self.channels,
                     rate=self.device_sample_rate,
                     input=True,
-                    frames_per_buffer=self.CHUNK,
+                    frames_per_buffer=self.chunk_size,
                     input_device_index=self.input_device_index,
                 )
                 if self.debug_mode:
@@ -135,6 +148,60 @@ class AudioInput:
                 self.audio_interface.terminate()
             return False
 
+    def lowpass_filter(self, signal, cutoff_freq, sample_rate):
+        """
+        Apply a low-pass Butterworth filter to prevent aliasing in the signal.
+
+        Args:
+            signal (np.ndarray): Input audio signal to filter
+            cutoff_freq (float): Cutoff frequency in Hz
+            sample_rate (float): Sampling rate of the input signal in Hz
+
+        Returns:
+            np.ndarray: Filtered audio signal
+
+        Notes:
+            - Uses a 5th order Butterworth filter
+            - Applies zero-phase filtering using filtfilt
+        """
+        # Calculate the Nyquist frequency (half the sample rate)
+        nyquist_rate = sample_rate / 2.0
+
+        # Normalize cutoff frequency to Nyquist rate (required by butter())
+        normal_cutoff = cutoff_freq / nyquist_rate
+
+        # Design the Butterworth filter
+        b, a = butter(5, normal_cutoff, btype='low', analog=False)
+
+        # Apply zero-phase filtering (forward and backward)
+        filtered_signal = filtfilt(b, a, signal)
+        return filtered_signal
+
+    def resample_audio(self, pcm_data, target_sample_rate, original_sample_rate):
+        """
+        Filter and resample audio data to a target sample rate.
+
+        Args:
+            pcm_data (np.ndarray): Input audio data
+            target_sample_rate (int): Desired output sample rate in Hz
+            original_sample_rate (int): Original sample rate of input in Hz
+
+        Returns:
+            np.ndarray: Resampled audio data
+
+        Notes:
+            - Applies anti-aliasing filter before resampling
+            - Uses polyphase filtering for high-quality resampling
+        """
+        if target_sample_rate < original_sample_rate:
+            # Downsampling with low-pass filter
+            pcm_filtered = self.lowpass_filter(pcm_data, target_sample_rate / 2, original_sample_rate)
+            resampled = resample_poly(pcm_filtered, target_sample_rate, original_sample_rate)
+        else:
+            # Upsampling without low-pass filter
+            resampled = resample_poly(pcm_data, target_sample_rate, original_sample_rate)
+        return resampled
+
     def read_chunk(self):
         """Read a chunk of audio data"""
         return self.stream.read(self.CHUNK, exception_on_overflow=False)

{realtimestt-0.3.92 → realtimestt-0.3.94}/RealtimeSTT/audio_recorder.py

@@ -120,7 +120,7 @@ class TranscriptionWorker:
                     data = self.conn.recv()
                     self.queue.put(data)
                 except Exception as e:
-                    logging.error(f"Error receiving data from connection: {e}")
+                    logging.error(f"Error receiving data from connection: {e}", exc_info=True)
             else:
                 time.sleep(TIME_SLEEP)
 
@@ -139,8 +139,13 @@ class TranscriptionWorker:
                 device_index=self.gpu_device_index,
                 download_root=self.download_root,
             )
+            # Create a short dummy audio array, for example 1 second of silence at 16 kHz
             if self.batch_size > 0:
                 model = BatchedInferencePipeline(model=model)
+
+            # Run a warm-up transcription
+            dummy_audio = np.zeros(16000, dtype=np.float32)
+            model.transcribe(dummy_audio, language="en", beam_size=1)
         except Exception as e:
             logging.exception(f"Error initializing main faster_whisper transcription model: {e}")
             raise
@@ -157,6 +162,7 @@ class TranscriptionWorker:
                 try:
                     audio, language = self.queue.get(timeout=0.1)
                     try:
+                        logging.debug(f"Transcribing audio with language {language}")
                         if self.batch_size > 0:
                             segments, info = model.transcribe(
                                 audio,
@@ -179,7 +185,7 @@ class TranscriptionWorker:
                         logging.debug(f"Final text detected with main model: {transcription}")
                         self.conn.send(('success', (transcription, info)))
                     except Exception as e:
-                        logging.error(f"General error in transcription: {e}")
+                        logging.error(f"General error in transcription: {e}", exc_info=True)
                         self.conn.send(('error', str(e)))
                 except queue.Empty:
                     continue
@@ -188,7 +194,7 @@ class TranscriptionWorker:
                     logging.debug("Transcription worker process finished due to KeyboardInterrupt")
                     break
                 except Exception as e:
-                    logging.error(f"General error in processing queue item: {e}")
+                    logging.error(f"General error in processing queue item: {e}", exc_info=True)
         finally:
             __builtins__['print'] = print  # Restore the original print function
             self.conn.close()
@@ -281,6 +287,7 @@ class AudioToTextRecorder:
                  buffer_size: int = BUFFER_SIZE,
                  sample_rate: int = SAMPLE_RATE,
                  initial_prompt: Optional[Union[str, Iterable[int]]] = None,
+                 initial_prompt_realtime: Optional[Union[str, Iterable[int]]] = None,
                  suppress_tokens: Optional[List[int]] = [-1],
                  print_transcription_time: bool = False,
                  early_transcription_on_silence: int = 0,
@@ -294,12 +301,12 @@ class AudioToTextRecorder:
 
         Args:
         - model (str, default="tiny"): Specifies the size of the transcription
-          model to use or the path to a converted model directory.
-                Valid options are 'tiny', 'tiny.en', 'base', 'base.en',
-                'small', 'small.en', 'medium', 'medium.en', 'large-v1',
-                'large-v2'.
-                If a specific size is provided, the model is downloaded
-                from the Hugging Face Hub.
+            model to use or the path to a converted model directory.
+            Valid options are 'tiny', 'tiny.en', 'base', 'base.en',
+            'small', 'small.en', 'medium', 'medium.en', 'large-v1',
+            'large-v2'.
+            If a specific size is provided, the model is downloaded
+            from the Hugging Face Hub.
         - download_root (str, default=None): Specifies the root path were the Whisper models 
           are downloaded to. When empty, the default is used. 
         - language (str, default=""): Language code for speech-to-text engine.
@@ -472,7 +479,9 @@ class AudioToTextRecorder:
             recording. Changing this will very probably functionality (as the
             WebRTC VAD model is very sensitive towards the sample rate).
         - initial_prompt (str or iterable of int, default=None): Initial
-            prompt to be fed to the transcription models.
+            prompt to be fed to the main transcription model.
+        - initial_prompt_realtime (str or iterable of int, default=None):
+            Initial prompt to be fed to the real-time transcription model.
         - suppress_tokens (list of int, default=[-1]): Tokens to be suppressed
             from the transcription output.
         - print_transcription_time (bool, default=False): Logs processing time
@@ -533,6 +542,8 @@ class AudioToTextRecorder:
         self.enable_realtime_transcription = enable_realtime_transcription
         self.use_main_model_for_realtime = use_main_model_for_realtime
         self.main_model_type = model
+        if not download_root:
+            download_root = None
         self.download_root = download_root
         self.realtime_model_type = realtime_model_type
         self.realtime_processing_pause = realtime_processing_pause
@@ -557,6 +568,8 @@ class AudioToTextRecorder:
         self.sample_rate = sample_rate
         self.recording_start_time = 0
         self.recording_stop_time = 0
+        self.last_recording_start_time = 0
+        self.last_recording_stop_time = 0
         self.wake_word_detect_time = 0
         self.silero_check_time = 0
         self.silero_working = False
@@ -580,9 +593,12 @@ class AudioToTextRecorder:
         self.stream = None
         self.start_recording_event = threading.Event()
         self.stop_recording_event = threading.Event()
+        self.backdate_stop_seconds = 0.0
+        self.backdate_resume_seconds = 0.0
         self.last_transcription_bytes = None
         self.last_transcription_bytes_b64 = None
         self.initial_prompt = initial_prompt
+        self.initial_prompt_realtime = initial_prompt_realtime
         self.suppress_tokens = suppress_tokens
         self.use_wake_words = wake_words or wakeword_backend in {'oww', 'openwakeword', 'openwakewords'}
         self.detected_language = None
@@ -697,7 +713,11 @@ class AudioToTextRecorder:
         if self.enable_realtime_transcription and not self.use_main_model_for_realtime:
             try:
                 logging.info("Initializing faster_whisper realtime "
-                             f"transcription model {self.realtime_model_type}"
+                             f"transcription model {self.realtime_model_type}, "
+                             f"default device: {self.device}, "
+                             f"compute type: {self.compute_type}, "
+                             f"device index: {self.gpu_device_index}, "
+                             f"download root: {self.download_root}"
                              )
                 self.realtime_model_type = faster_whisper.WhisperModel(
                     model_size_or_path=self.realtime_model_type,
@@ -708,7 +728,8 @@ class AudioToTextRecorder:
                 )
                 if self.realtime_batch_size > 0:
                     self.realtime_model_type = BatchedInferencePipeline(model=self.realtime_model_type)
-
+                dummy_audio = np.zeros(16000, dtype=np.float32)
+                self.realtime_model_type.transcribe(dummy_audio, language="en", beam_size=1)
             except Exception as e:
                 logging.exception("Error initializing faster_whisper "
                                   f"realtime transcription model: {e}"
@@ -844,6 +865,7 @@ class AudioToTextRecorder:
                        0.3)
         )
         self.frames = []
+        self.last_frames = []
 
         # Recording control flags
         self.is_recording = False
@@ -909,7 +931,7 @@ class AudioToTextRecorder:
                 logging.info("KeyboardInterrupt in read from stdout detected, exiting...")
                 break
             except Exception as e:
-                logging.error(f"Unexpected error in read from stdout: {e}")
+                logging.error(f"Unexpected error in read from stdout: {e}", exc_info=True)
                 logging.error(traceback.format_exc())  # Log the full traceback here
                 break 
             time.sleep(0.1)
@@ -1058,7 +1080,7 @@ class AudioToTextRecorder:
                     return stream
 
                 except Exception as e:
-                    logging.error(f"Microphone connection failed: {e}. Retrying...")
+                    logging.error(f"Microphone connection failed: {e}. Retrying...", exc_info=True)
                     input_device_index = None
                     time.sleep(3)  # Wait before retrying
                     continue
@@ -1174,7 +1196,7 @@ class AudioToTextRecorder:
                     if e.errno == pyaudio.paInputOverflowed:
                         logging.warning("Input overflowed. Frame dropped.")
                     else:
-                        logging.error(f"OSError during recording: {e}")
+                        logging.error(f"OSError during recording: {e}", exc_info=True)
                         # Attempt to reinitialize the stream
                         logging.error("Attempting to reinitialize the audio stream...")
 
@@ -1292,13 +1314,60 @@ class AudioToTextRecorder:
                     if (self.stop_recording_event.wait(timeout=0.02)):
                         break
 
-            # Convert recorded frames to the appropriate audio format.
-            audio_array = np.frombuffer(b''.join(self.frames), dtype=np.int16)
-            self.audio = audio_array.astype(np.float32) / INT16_MAX_ABS_VALUE
+            frames = self.frames
+            if len(frames) == 0:
+                frames = self.last_frames
+
+            # Calculate samples needed for backdating resume
+            samples_to_keep = int(self.sample_rate * self.backdate_resume_seconds)
+
+            # First convert all current frames to audio array
+            full_audio_array = np.frombuffer(b''.join(frames), dtype=np.int16)
+            full_audio = full_audio_array.astype(np.float32) / INT16_MAX_ABS_VALUE
+
+            # Calculate how many samples we need to keep for backdating resume
+            if samples_to_keep > 0:
+                samples_to_keep = min(samples_to_keep, len(full_audio))
+                # Keep the last N samples for backdating resume
+                frames_to_read_audio = full_audio[-samples_to_keep:]
+
+                # Convert the audio back to int16 bytes for frames
+                frames_to_read_int16 = (frames_to_read_audio * INT16_MAX_ABS_VALUE).astype(np.int16)
+                frame_bytes = frames_to_read_int16.tobytes()
+
+                # Split into appropriate frame sizes (assuming standard frame size)
+                FRAME_SIZE = 2048  # Typical frame size
+                frames_to_read = []
+                for i in range(0, len(frame_bytes), FRAME_SIZE):
+                    frame = frame_bytes[i:i + FRAME_SIZE]
+                    if frame:  # Only add non-empty frames
+                        frames_to_read.append(frame)
+            else:
+                frames_to_read = []
+
+            # Process backdate stop seconds
+            samples_to_remove = int(self.sample_rate * self.backdate_stop_seconds)
+
+            if samples_to_remove > 0:
+                if samples_to_remove < len(full_audio):
+                    self.audio = full_audio[:-samples_to_remove]
+                    logging.debug(f"Removed {samples_to_remove} samples "
+                        f"({samples_to_remove/self.sample_rate:.3f}s) from end of audio")
+                else:
+                    self.audio = np.array([], dtype=np.float32)
+                    logging.debug("Cleared audio (samples_to_remove >= audio length)")
+            else:
+                self.audio = full_audio
+                logging.debug(f"No samples removed, final audio length: {len(self.audio)}")
+
             self.frames.clear()
+            self.last_frames.clear()
+            self.frames.extend(frames_to_read)
+
+            # Reset backdating parameters
+            self.backdate_stop_seconds = 0.0
+            self.backdate_resume_seconds = 0.0
 
-            # Reset recording-related timestamps
-            self.recording_stop_time = 0
             self.listen_start = 0
 
             self._set_state("inactive")
@@ -1372,7 +1441,7 @@ class AudioToTextRecorder:
                     logging.error(f"Transcription error: {result}")
                     raise Exception(result)
             except Exception as e:
-                logging.error(f"Error during transcription: {str(e)}")
+                logging.error(f"Error during transcription: {str(e)}", exc_info=True)
                 raise e
 
     def _process_wakeword(self, data):
@@ -1460,7 +1529,16 @@ class AudioToTextRecorder:
             return self.transcribe()
 
 
-    def start(self):
+    def format_number(self, num):
+        # Convert the number to a string
+        num_str = f"{num:.10f}"  # Ensure precision is sufficient
+        # Split the number into integer and decimal parts
+        integer_part, decimal_part = num_str.split('.')
+        # Take the last two digits of the integer part and the first two digits of the decimal part
+        result = f"{integer_part[-2:]}.{decimal_part[:2]}"
+        return result
+
+    def start(self, frames = None):
         """
         Starts recording audio directly without waiting for voice activity.
         """
@@ -1482,7 +1560,10 @@ class AudioToTextRecorder:
         self.wakeword_detected = False
         self.wake_word_detect_time = 0
         self.frames = []
+        if frames:
+            self.frames = frames
         self.is_recording = True
+
         self.recording_start_time = time.time()
         self.is_silero_speech_active = False
         self.is_webrtc_speech_active = False
@@ -1494,9 +1575,19 @@ class AudioToTextRecorder:
 
         return self
 
-    def stop(self):
+    def stop(self,
+             backdate_stop_seconds: float = 0.0,
+             backdate_resume_seconds: float = 0.0,
+        ):
         """
         Stops recording audio.
+
+        Args:
+        - backdate_stop_seconds (float, default="0.0"): Specifies the number of
+            seconds to backdate the stop time. This is useful when the stop
+            command is issued after the actual stop time.
+        - backdate_resume_seconds (float, default="0.0"): Specifies the number
+            of seconds to backdate the time relistening is initiated.
         """
 
         # Ensure there's a minimum interval
@@ -1509,6 +1600,9 @@ class AudioToTextRecorder:
             return self
 
         logging.info("recording stopped")
+        self.last_frames = copy.deepcopy(self.frames)
+        self.backdate_stop_seconds = backdate_stop_seconds
+        self.backdate_resume_seconds = backdate_resume_seconds
         self.is_recording = False
         self.recording_stop_time = time.time()
         self.is_silero_speech_active = False
@@ -1517,6 +1611,9 @@ class AudioToTextRecorder:
         self.start_recording_event.clear()
         self.stop_recording_event.set()
 
+        self.last_recording_start_time = self.recording_start_time
+        self.last_recording_stop_time = self.recording_stop_time
+
         if self.on_recording_stop:
             self.on_recording_stop()
 
@@ -1591,7 +1688,6 @@ class AudioToTextRecorder:
                 return
 
             print("\033[91mRealtimeSTT shutting down\033[0m")
-            # logging.debug("RealtimeSTT shutting down")
 
             # Force wait_audio() and text() to exit
             self.is_shut_down = True
@@ -1720,7 +1816,7 @@ class AudioToTextRecorder:
                             data = self.audio_queue.get()
 
                 except BrokenPipeError:
-                    logging.error("BrokenPipeError _recording_worker")
+                    logging.error("BrokenPipeError _recording_worker", exc_info=True)
                     self.is_running = False
                     break
 
@@ -1797,11 +1893,11 @@ class AudioToTextRecorder:
 
                         except struct.error:
                             logging.error("Error unpacking audio data "
-                                        "for wake word processing.")
+                                        "for wake word processing.", exc_info=True)
                             continue
 
                         except Exception as e:
-                            logging.error(f"Wake word processing error: {e}")
+                            logging.error(f"Wake word processing error: {e}", exc_info=True)
                             continue
 
                         if self.use_extended_logging:
@@ -2028,7 +2124,7 @@ class AudioToTextRecorder:
         except Exception as e:
             logging.debug('Debug: Caught exception in main try block')
             if not self.interrupt_stop_event.is_set():
-                logging.error(f"Unhandled exeption in _recording_worker: {e}")
+                logging.error(f"Unhandled exeption in _recording_worker: {e}", exc_info=True)
                 raise
 
         if self.use_extended_logging:
@@ -2095,7 +2191,7 @@ class AudioToTextRecorder:
                                     logging.warning("Realtime transcription timed out")
                                     continue
                             except Exception as e:
-                                logging.error(f"Error in realtime transcription: {str(e)}")
+                                logging.error(f"Error in realtime transcription: {str(e)}", exc_info=True)
                                 continue
                     else:
                         # Perform transcription and assemble the text
@@ -2104,7 +2200,7 @@ class AudioToTextRecorder:
                                 audio_array,
                                 language=self.language if self.language else None,
                                 beam_size=self.beam_size_realtime,
-                                initial_prompt=self.initial_prompt,
+                                initial_prompt=self.initial_prompt_realtime,
                                 suppress_tokens=self.suppress_tokens,
                                 batch_size=self.realtime_batch_size
                             )
@@ -2113,7 +2209,7 @@ class AudioToTextRecorder:
                                 audio_array,
                                 language=self.language if self.language else None,
                                 beam_size=self.beam_size_realtime,
-                                initial_prompt=self.initial_prompt,
+                                initial_prompt=self.initial_prompt_realtime,
                                 suppress_tokens=self.suppress_tokens
                             )
 
@@ -2129,7 +2225,6 @@ class AudioToTextRecorder:
                     if self.is_recording and time.time() - \
                             self.recording_start_time > self.init_realtime_after_seconds:
 
-                        # logging.debug('Starting realtime transcription')
                         self.realtime_transcription_text = realtime_text
                         self.realtime_transcription_text = \
                             self.realtime_transcription_text.strip()
@@ -2209,7 +2304,7 @@ class AudioToTextRecorder:
                     time.sleep(TIME_SLEEP)
 
         except Exception as e:
-            logging.error(f"Unhandled exeption in _realtime_worker: {e}")
+            logging.error(f"Unhandled exeption in _realtime_worker: {e}", exc_info=True)
             raise
 
     def _is_silero_speech(self, chunk):

{realtimestt-0.3.92 → realtimestt-0.3.94}/RealtimeSTT/audio_recorder_client.py

@@ -28,6 +28,7 @@ DEFAULT_DATA_URL = "ws://127.0.0.1:8012"
 INIT_MODEL_TRANSCRIPTION = "tiny"
 INIT_MODEL_TRANSCRIPTION_REALTIME = "tiny"
 INIT_REALTIME_PROCESSING_PAUSE = 0.2
+INIT_REALTIME_INITIAL_PAUSE = 0.2
 INIT_SILERO_SENSITIVITY = 0.4
 INIT_WEBRTC_SENSITIVITY = 3
 INIT_POST_SPEECH_SILENCE_DURATION = 0.6
@@ -68,6 +69,7 @@ class AudioToTextRecorderClient:
 
     def __init__(self,
                  model: str = INIT_MODEL_TRANSCRIPTION,
+                 download_root: str = None, 
                  language: str = "",
                  compute_type: str = "default",
                  input_device_index: int = None,
@@ -81,14 +83,17 @@ class AudioToTextRecorderClient:
                  use_microphone=True,
                  spinner=True,
                  level=logging.WARNING,
+                 batch_size: int = 16,
 
                  # Realtime transcription parameters
                  enable_realtime_transcription=False,
                  use_main_model_for_realtime=False,
                  realtime_model_type=INIT_MODEL_TRANSCRIPTION_REALTIME,
                  realtime_processing_pause=INIT_REALTIME_PROCESSING_PAUSE,
+                 init_realtime_after_seconds=INIT_REALTIME_INITIAL_PAUSE,
                  on_realtime_transcription_update=None,
                  on_realtime_transcription_stabilized=None,
+                 realtime_batch_size: int = 16,
 
                  # Voice activation parameters
                  silero_sensitivity: float = INIT_SILERO_SENSITIVITY,
@@ -133,6 +138,7 @@ class AudioToTextRecorderClient:
                  buffer_size: int = BUFFER_SIZE,
                  sample_rate: int = SAMPLE_RATE,
                  initial_prompt: Optional[Union[str, Iterable[int]]] = None,
+                 initial_prompt_realtime: Optional[Union[str, Iterable[int]]] = None,
                  suppress_tokens: Optional[List[int]] = [-1],
                  print_transcription_time: bool = False,
                  early_transcription_on_silence: int = 0,
@@ -162,10 +168,14 @@ class AudioToTextRecorderClient:
         self.use_microphone = use_microphone
         self.spinner = spinner
         self.level = level
+        self.batch_size = batch_size
+        self.init_realtime_after_seconds = init_realtime_after_seconds
+        self.realtime_batch_size = realtime_batch_size
 
         # Real-time transcription parameters
         self.enable_realtime_transcription = enable_realtime_transcription
         self.use_main_model_for_realtime = use_main_model_for_realtime
+        self.download_root = download_root
         self.realtime_model_type = realtime_model_type
         self.realtime_processing_pause = realtime_processing_pause
         self.on_realtime_transcription_update = on_realtime_transcription_update
@@ -204,6 +214,7 @@ class AudioToTextRecorderClient:
         self.buffer_size = buffer_size
         self.sample_rate = sample_rate
         self.initial_prompt = initial_prompt
+        self.initial_prompt_realtime = initial_prompt_realtime
         self.suppress_tokens = suppress_tokens
         self.print_transcription_time = print_transcription_time
         self.early_transcription_on_silence = early_transcription_on_silence
@@ -376,6 +387,43 @@ class AudioToTextRecorderClient:
             args += ['--model', self.model]
         if self.realtime_model_type:
             args += ['--realtime_model_type', self.realtime_model_type]
+        if self.download_root:
+            args += ['--root', self.download_root]
+        if self.batch_size is not None:
+            args += ['--batch', str(self.batch_size)]
+        if self.realtime_batch_size is not None:
+            args += ['--realtime_batch_size', str(self.realtime_batch_size)]
+        if self.init_realtime_after_seconds is not None:
+            args += ['--init_realtime_after_seconds', str(self.init_realtime_after_seconds)]
+        if self.initial_prompt_realtime:
+            sanitized_prompt = self.initial_prompt_realtime.replace("\n", "\\n")
+            args += ['--initial_prompt_realtime', sanitized_prompt]
+
+        # if self.compute_type:
+        #     args += ['--compute_type', self.compute_type]
+        # if self.input_device_index is not None:
+        #     args += ['--input_device_index', str(self.input_device_index)]
+        # if self.gpu_device_index is not None:
+        #     args += ['--gpu_device_index', str(self.gpu_device_index)]
+        # if self.device:
+        #     args += ['--device', self.device]
+        # if self.spinner:
+        #     args.append('--spinner')  # flag, no need for True/False
+        # if self.enable_realtime_transcription:
+        #     args.append('--enable_realtime_transcription')  # flag, no need for True/False
+        # if self.handle_buffer_overflow:
+        #     args.append('--handle_buffer_overflow')  # flag, no need for True/False
+        # if self.suppress_tokens:
+        #     args += ['--suppress_tokens', str(self.suppress_tokens)]
+        # if self.print_transcription_time:
+        #     args.append('--print_transcription_time')  # flag, no need for True/False
+        # if self.allowed_latency_limit is not None:
+        #     args += ['--allowed_latency_limit', str(self.allowed_latency_limit)]
+        # if self.no_log_file:
+        #     args.append('--no_log_file')  # flag, no need for True
+        # if self.debug_mode:
+        #     args.append('--debug')  # flag, no need for True/False
+            
         if self.language:
             args += ['--language', self.language]
         if self.silero_sensitivity is not None:

{realtimestt-0.3.92 → realtimestt-0.3.94}/RealtimeSTT.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: RealtimeSTT
-Version: 0.3.92
+Version: 0.3.94
 Summary: A fast Voice Activity Detection and Transcription System
 Home-page: https://github.com/KoljaB/RealTimeSTT
 Author: Kolja Beigel
@@ -13,15 +13,15 @@ Requires-Python: >=3.6
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: PyAudio==0.2.14
-Requires-Dist: faster-whisper==1.1.0
+Requires-Dist: faster-whisper==1.1.1
 Requires-Dist: pvporcupine==1.9.5
 Requires-Dist: webrtcvad-wheels==2.0.14
 Requires-Dist: halo==0.0.31
 Requires-Dist: torch
 Requires-Dist: torchaudio
-Requires-Dist: scipy==1.14.1
+Requires-Dist: scipy==1.15.1
 Requires-Dist: openwakeword>=0.4.0
-Requires-Dist: websockets==14.1
+Requires-Dist: websockets==14.2
 Requires-Dist: websocket-client==1.8.0
 Requires-Dist: numpy<2.0.0
 

{realtimestt-0.3.92 → realtimestt-0.3.94}/RealtimeSTT.egg-info/requires.txt

@@ -1,12 +1,12 @@
 PyAudio==0.2.14
-faster-whisper==1.1.0
+faster-whisper==1.1.1
 pvporcupine==1.9.5
 webrtcvad-wheels==2.0.14
 halo==0.0.31
 torch
 torchaudio
-scipy==1.14.1
+scipy==1.15.1
 openwakeword>=0.4.0
-websockets==14.1
+websockets==14.2
 websocket-client==1.8.0
 numpy<2.0.0

{realtimestt-0.3.92 → realtimestt-0.3.94}/RealtimeSTT_server/stt_server.py

@@ -27,6 +27,8 @@ stt-server [OPTIONS]
     - `-D, --debug`: Enable debug logging.
     - `-W, --write`: Save audio to WAV file.
     - `-s, --silence_timing`: Enable dynamic silence duration for sentence detection; default True. 
+    - `-b, --batch, --batch_size`: Batch size for inference; default 16.
+    - `--root, --download_root`: Specifies the root path were the Whisper models are downloaded to.
     - `--silero_sensitivity`: Silero VAD sensitivity (0-1); default 0.05.
     - `--silero_use_onnx`: Use Silero ONNX model; default False.
     - `--webrtc_sensitivity`: WebRTC VAD sensitivity (0-3); default 3.
@@ -38,7 +40,10 @@ stt-server [OPTIONS]
     - `--early_transcription_on_silence`: Start transcription after silence in seconds; default 0.2.
     - `--beam_size`: Beam size for main model; default 5.
     - `--beam_size_realtime`: Beam size for real-time model; default 3.
-    - `--initial_prompt`: Initial transcription guidance prompt.
+    - `--init_realtime_after_seconds`: Initial waiting time for realtime transcription; default 0.2.
+    - `--realtime_batch_size`: Batch size for the real-time transcription model; default 16.
+    - `--initial_prompt`: Initial main transcription guidance prompt.
+    - `--initial_prompt_realtime`: Initial realtime transcription guidance prompt.
     - `--end_of_sentence_detection_pause`: Silence duration for sentence end detection; default 0.45.
     - `--unknown_sentence_detection_pause`: Pause duration for incomplete sentence detection; default 0.7.
     - `--mid_sentence_detection_pause`: Pause for mid-sentence break; default 2.0.
@@ -52,6 +57,14 @@ stt-server [OPTIONS]
     - `--use_main_model_for_realtime`: Use main model for real-time transcription.
     - `--use_extended_logging`: Enable extensive log messages.
     - `--logchunks`: Log incoming audio chunks.
+    - `--compute_type`: Type of computation to use.
+    - `--input_device_index`: Index of the audio input device.
+    - `--gpu_device_index`: Index of the GPU device.
+    - `--device`: Device to use for computation.
+    - `--handle_buffer_overflow`: Handle buffer overflow during transcription.
+    - `--suppress_tokens`: Suppress tokens during transcription.
+    - `--allowed_latency_limit`: Allowed latency limit for real-time transcription.
+
 
 ### WebSocket Interface:
 The server supports two WebSocket connections:
@@ -364,7 +377,7 @@ def parse_arguments():
     parser.add_argument('-l', '--lang', '--language', type=str, default='en',
                 help='Language code for the STT model to transcribe in a specific language. Leave this empty for auto-detection based on input audio. Default is en. List of supported language codes: https://github.com/openai/whisper/blob/main/whisper/tokenizer.py#L11-L110')
 
-    parser.add_argument('-i', '--input-device', '--input_device_index', type=int, default=1,
+    parser.add_argument('-i', '--input-device', '--input-device-index', type=int, default=1,
                     help='Index of the audio input device to use. Use this option to specify a particular microphone or audio input device based on your system. Default is 1.')
 
     parser.add_argument('-c', '--control', '--control_port', type=int, default=8011,
@@ -378,12 +391,23 @@ def parse_arguments():
 
     parser.add_argument('-D', '--debug', action='store_true', help='Enable debug logging for detailed server operations')
 
-    parser.add_argument("-W", "--write", metavar="FILE",
-                        help="Save received audio to a WAV file")
+    parser.add_argument('-W', '--write', metavar='FILE', help='Save received audio to a WAV file')
+    
+    parser.add_argument('-b', '--batch', '--batch_size', type=int, default=16, help='Batch size for inference. This parameter controls the number of audio chunks processed in parallel during transcription. Default is 16.')
+
+    parser.add_argument('--root', '--download_root', type=str,default=None, help='Specifies the root path where the Whisper models are downloaded to. Default is None.')
 
     parser.add_argument('-s', '--silence_timing', action='store_true', default=True,
                     help='Enable dynamic adjustment of silence duration for sentence detection. Adjusts post-speech silence duration based on detected sentence structure and punctuation. Default is False.')
 
+    parser.add_argument('--init_realtime_after_seconds', type=float, default=0.2,
+                        help='The initial waiting time in seconds before real-time transcription starts. This delay helps prevent false positives at the beginning of a session. Default is 0.2 seconds.')  
+    
+    parser.add_argument('--realtime_batch_size', type=int, default=16,
+                        help='Batch size for the real-time transcription model. This parameter controls the number of audio chunks processed in parallel during real-time transcription. Default is 16.')
+    
+    parser.add_argument('--initial_prompt_realtime', type=str, default="", help='Initial prompt that guides the real-time transcription model to produce transcriptions in a particular style or format.')
+
     parser.add_argument('--silero_sensitivity', type=float, default=0.05,
                         help='Sensitivity level for Silero Voice Activity Detection (VAD), with a range from 0 to 1. Lower values make the model less sensitive, useful for noisy environments. Default is 0.05.')
 
@@ -457,6 +481,23 @@ def parse_arguments():
     parser.add_argument('--use_extended_logging', action='store_true',
                         help='Writes extensive log messages for the recording worker, that processes the audio chunks.')
 
+    parser.add_argument('--compute_type', type=str, default='default',
+                        help='Type of computation to use. See https://opennmt.net/CTranslate2/quantization.html')
+
+    parser.add_argument('--gpu_device_index', type=int, default=0,
+                        help='Index of the GPU device to use. Default is None.')
+    
+    parser.add_argument('--device', type=str, default='cuda',
+                        help='Device for model to use. Can either be "cuda" or "cpu". Default is cuda.')
+    
+    parser.add_argument('--handle_buffer_overflow', action='store_true',
+                        help='Handle buffer overflow during transcription. Default is False.')
+
+    parser.add_argument('--suppress_tokens', type=int, default=[-1], nargs='*', help='Suppress tokens during transcription. Default is [-1].')
+
+    parser.add_argument('--allowed_latency_limit', type=int, default=100,
+                        help='Maximal amount of chunks that can be unprocessed in queue before discarding chunks.. Default is 100.')
+
     parser.add_argument('--logchunks', action='store_true', help='Enable logging of incoming audio chunks (periods)')
 
     # Parse arguments
@@ -479,6 +520,9 @@ def parse_arguments():
     if args.initial_prompt:
         args.initial_prompt = args.initial_prompt.replace("\\n", "\n")
 
+    if args.initial_prompt_realtime:
+        args.initial_prompt_realtime = args.initial_prompt_realtime.replace("\\n", "\n")
+
     return args
 
 def _recorder_thread(loop):
@@ -534,7 +578,7 @@ def decode_and_resample(
 
     return resampled_audio.astype(np.int16).tobytes()
 
-async def control_handler(websocket, path):
+async def control_handler(websocket):
     debug_print(f"New control connection from {websocket.remote_address}")
     print(f"{bcolors.OKGREEN}Control client connected{bcolors.ENDC}")
     global recorder
@@ -629,7 +673,7 @@ async def control_handler(websocket, path):
     finally:
         control_connections.remove(websocket)
 
-async def data_handler(websocket, path):
+async def data_handler(websocket):
     global writechunks, wav_file
     print(f"{bcolors.OKGREEN}Data client connected{bcolors.ENDC}")
     data_connections.add(websocket)
@@ -700,8 +744,13 @@ async def main_async():
 
     recorder_config = {
         'model': args.model,
+        'download_root': args.root,
         'realtime_model_type': args.rt_model,
         'language': args.lang,
+        'batch_size': args.batch,
+        'init_realtime_after_seconds': args.init_realtime_after_seconds,
+        'realtime_batch_size': args.realtime_batch_size,
+        'initial_prompt_realtime': args.initial_prompt_realtime,
         'input_device_index': args.input_device,
         'silero_sensitivity': args.silero_sensitivity,
         'silero_use_onnx': args.silero_use_onnx,
@@ -740,6 +789,12 @@ async def main_async():
         'no_log_file': True,  # Disable logging to file
         'use_extended_logging': args.use_extended_logging,
         'level': loglevel,
+        'compute_type': args.compute_type,
+        'gpu_device_index': args.gpu_device_index,
+        'device': args.device,
+        'handle_buffer_overflow': args.handle_buffer_overflow,
+        'suppress_tokens': args.suppress_tokens,
+        'allowed_latency_limit': args.allowed_latency_limit,
     }
 
     try:

{realtimestt-0.3.92 → realtimestt-0.3.94}/setup.py

@@ -9,7 +9,7 @@ with open('requirements.txt') as f:
 
 setuptools.setup(
     name="RealtimeSTT",
-    version="0.3.92",
+    version="0.3.94",
     author="Kolja Beigel",
     author_email="kolja.beigel@web.de",
     description="A fast Voice Activity Detection and Transcription System",