Streaming speech recognition with Google Speech-to-Text is leading to improperly timestamped transcripts

Question

My Problem:

The web app I'm building relies on real-time transcription of a user's voice along with timestamps for when each word begins and ends.

Google's Speech-to-Text API has a limit of 4 minutes for streaming requests but I want users to be able to run their mic's for as long as 30 minutes if they so choose.

Thankfully, Google provides its own code examples for how to make successive requests to their Speech-to-Text API in a way that mimics endless streaming speech recognition.

I've adapted their Python infinite streaming example for my purposes (see below for my code). The timestamps provided by Google are pretty accurate but the issue is that when I exceed the streaming limit (4 minutes) and a new request is made, the timestamped transcript returned by Google's API from the new request is off by as much as 5 seconds or more.

Below is an example of the output when I adjust the streaming limit to 10 seconds (so a new request to Google's Speech-to-Text API begins every 10 seconds).

The timestamp you see printed next to each transcribed response (the 'corrected_time' in the code) is the timestamp for the end of the transcribed line, not the beginning. These timestamps are accurate for the first request but are off by ~4 seconds in the second request and ~9 seconds in the third request.

In a Nutshell, I want to make sure that when the streaming limit is exceeded and a new request is made, the timestamps returned by Google for that new request are adjusted accurately.

My Code:

To help you understand what's going on, I would recommend running it on your machine (only takes a couple of minutes to get working if you have a Google Cloud service account).

I've included more detail on my current diagnosis below the code.

#!/usr/bin/env python

"""Google Cloud Speech API sample application using the streaming API.

NOTE: This module requires the dependencies `pyaudio`.
To install using pip:

    pip install pyaudio

Example usage:
    python THIS_FILENAME.py
"""

# [START speech_transcribe_infinite_streaming]
import os
import re
import sys
import time

from google.cloud import speech
import pyaudio
from six.moves import queue

# Audio recording parameters
STREAMING_LIMIT = 20000  # 20 seconds (originally 4 mins but shortened for testing purposes)
SAMPLE_RATE = 16000
CHUNK_SIZE = int(SAMPLE_RATE / 10)  # 100ms

# Environment Variable set for Google Credentials. Put the json service account
# key in the root directory
os.environ['GOOGLE_APPLICATION_CREDENTIALS'] = 'YOUR_SERVICE_ACCOUNT_KEY.json'

def get_current_time():
    """Return Current Time in MS."""

    return int(round(time.time() * 1000))


class ResumableMicrophoneStream:
    """Opens a recording stream as a generator yielding the audio chunks."""

    def __init__(self, rate, chunk_size):
        self._rate = rate
        self.chunk_size = chunk_size
        self._num_channels = 1
        self._buff = queue.Queue()
        self.closed = True
        self.start_time = get_current_time()
        self.restart_counter = 0
        self.audio_input = []
        self.last_audio_input = []
        self.result_end_time = 0
        self.is_final_end_time = 0
        self.final_request_end_time = 0
        self.bridging_offset = 0
        self.last_transcript_was_final = False
        self.new_stream = True
        self._audio_interface = pyaudio.PyAudio()
        self._audio_stream = self._audio_interface.open(
            format=pyaudio.paInt16,
            channels=self._num_channels,
            rate=self._rate,
            input=True,
            frames_per_buffer=self.chunk_size,
            # Run the audio stream asynchronously to fill the buffer object.
            # This is necessary so that the input device's buffer doesn't
            # overflow while the calling thread makes network requests, etc.
            stream_callback=self._fill_buffer,
        )

    def __enter__(self):

        self.closed = False
        return self

    def __exit__(self, type, value, traceback):

        self._audio_stream.stop_stream()
        self._audio_stream.close()
        self.closed = True
        # Signal the generator to terminate so that the client's
        # streaming_recognize method will not block the process termination.
        self._buff.put(None)
        self._audio_interface.terminate()

    def _fill_buffer(self, in_data, *args, **kwargs):
        """Continuously collect data from the audio stream, into the buffer."""

        self._buff.put(in_data)
        return None, pyaudio.paContinue

    def generator(self):
        """Stream Audio from microphone to API and to local buffer"""

        while not self.closed:
            data = []
            
            """
            THE BELOW 'IF' STATEMENT IS WHERE THE ERROR IS LIKELY OCCURRING
            This statement runs when the streaming limit is hit and a new request is made.
            """
            if self.new_stream and self.last_audio_input:

                chunk_time = STREAMING_LIMIT / len(self.last_audio_input)

                if chunk_time != 0:

                    if self.bridging_offset < 0:
                        self.bridging_offset = 0

                    if self.bridging_offset > self.final_request_end_time:
                        self.bridging_offset = self.final_request_end_time

                    chunks_from_ms = round(
                        (self.final_request_end_time - self.bridging_offset)
                        / chunk_time
                    )

                    self.bridging_offset = round(
                        (len(self.last_audio_input) - chunks_from_ms) * chunk_time
                    )

                    for i in range(chunks_from_ms, len(self.last_audio_input)):
                        data.append(self.last_audio_input[i])

                self.new_stream = False

            # Use a blocking get() to ensure there's at least one chunk of
            # data, and stop iteration if the chunk is None, indicating the
            # end of the audio stream.
            chunk = self._buff.get()
            self.audio_input.append(chunk)

            if chunk is None:
                return
            data.append(chunk)
            # Now consume whatever other data's still buffered.
            while True:
                try:
                    chunk = self._buff.get(block=False)

                    if chunk is None:
                        return
                    data.append(chunk)
                    self.audio_input.append(chunk)

                except queue.Empty:
                    break

            yield b"".join(data)


def listen_print_loop(responses, stream):
    """Iterates through server responses and prints them.

    The responses passed is a generator that will block until a response
    is provided by the server.

    Each response may contain multiple results, and each result may contain
    multiple alternatives; Here we print only the transcription for the top 
    alternative of the top result.

    In this case, responses are provided for interim results as well. If the
    response is an interim one, print a line feed at the end of it, to allow
    the next result to overwrite it, until the response is a final one. For the
    final one, print a newline to preserve the finalized transcription.
    """

    for response in responses:

        if get_current_time() - stream.start_time > STREAMING_LIMIT:
            stream.start_time = get_current_time()
            break

        if not response.results:
            continue

        result = response.results[0]

        if not result.alternatives:
            continue

        transcript = result.alternatives[0].transcript

        result_seconds = 0
        result_micros = 0

        if result.result_end_time.seconds:
            result_seconds = result.result_end_time.seconds

        if result.result_end_time.microseconds:
            result_micros = result.result_end_time.microseconds

        stream.result_end_time = int((result_seconds * 1000) + (result_micros / 1000))

        corrected_time = (
            stream.result_end_time
            - stream.bridging_offset
            + (STREAMING_LIMIT * stream.restart_counter)
        )
        # Display interim results, but with a carriage return at the end of the
        # line, so subsequent lines will overwrite them.

        if result.is_final:

            sys.stdout.write("FINAL RESULT @ ")
            sys.stdout.write(str(corrected_time/1000) + ": " + transcript + "\n")

            stream.is_final_end_time = stream.result_end_time
            stream.last_transcript_was_final = True

            # Exit recognition if any of the transcribed phrases could be
            # one of our keywords.
            if re.search(r"\b(exit|quit)\b", transcript, re.I):
                sys.stdout.write("Exiting...\n")
                stream.closed = True
                break

        else:
            sys.stdout.write("INTERIM RESULT @ ")
            sys.stdout.write(str(corrected_time/1000) + ": " + transcript + "\r")

            stream.last_transcript_was_final = False


def main():
    """start bidirectional streaming from microphone input to speech API"""

    client = speech.SpeechClient()
    config = speech.RecognitionConfig(
        encoding=speech.RecognitionConfig.AudioEncoding.LINEAR16,
        sample_rate_hertz=SAMPLE_RATE,
        language_code="en-US",
        max_alternatives=1,
    )

    streaming_config = speech.StreamingRecognitionConfig(
        config=config, interim_results=True
    )

    mic_manager = ResumableMicrophoneStream(SAMPLE_RATE, CHUNK_SIZE)
    print(mic_manager.chunk_size)
    sys.stdout.write('\nListening, say "Quit" or "Exit" to stop.\n\n')
    sys.stdout.write("End (ms)       Transcript Results/Status\n")
    sys.stdout.write("=====================================================\n")

    with mic_manager as stream:

        while not stream.closed:
            sys.stdout.write(
                "\n" + str(STREAMING_LIMIT * stream.restart_counter) + ": NEW REQUEST\n"
            )

            stream.audio_input = []
            audio_generator = stream.generator()

            requests = (
                speech.StreamingRecognizeRequest(audio_content=content)
                for content in audio_generator
            )

            responses = client.streaming_recognize(streaming_config, requests)

            # Now, put the transcription responses to use.
            listen_print_loop(responses, stream)

            if stream.result_end_time > 0:
                stream.final_request_end_time = stream.is_final_end_time
            stream.result_end_time = 0
            stream.last_audio_input = []
            stream.last_audio_input = stream.audio_input
            stream.audio_input = []
            stream.restart_counter = stream.restart_counter + 1

            if not stream.last_transcript_was_final:
                sys.stdout.write("\n")
            stream.new_stream = True


if __name__ == "__main__":

    main()

# [END speech_transcribe_infinite_streaming]

My Current Diagnosis

The 'corrected_time' is not being set correctly when new requests are made. This is due to the 'bridging_offset' not being set correctly. So what we need to look at is the 'generator()' method in the 'ResumableMicrophoneStream' class.

In the 'generator()' method, there is an 'if' statement which is run when the streaming limit is hit and a new request is made

if self.new_stream and self.last_audio_input:

Its purpose appears to be to take any lingering audio data that wasn't finished being transcribed before the streaming limit was hit and add it to the buffer before any new audio chunks so that it's transcribed in the new request.

It is also the responsibility of this 'if' statement to set the 'bridging offset' but I'm not entirely sure what this offset represents. All I know is that however it is being set, it is not being set accurately.

Answer 1

Time offset values show the beginning and the end of each spoken word that is recognized in the supplied audio. A time offset value represents the amount of time that has elapsed from the beginning of the audio, in increments of 100ms.

This tells us that the offset you are receiving for each of the timestamps that you are running within your project will always make the timestamps from start to finish. That would be my guess as to why it’s causing your application problems.