Speech api: Fill audio buffer in a separate thread. (#527)

This is to avoid timing issues where the request thread doesn't poll the
generator fast enough to consume all the incoming audio data from the input
device. In that case, the audio device buffer overflows, leading to lost data
and exceptions and other nastiness.

Address #515

Author: jerjou

speech/grpc/requirements.txt

@@ -2,3 +2,4 @@ gcloud==0.18.1
 grpcio==1.0.0
 PyAudio==0.2.9
 grpc-google-cloud-speech-v1beta1==1.0.1
+six==1.10.0

speech/grpc/transcribe_streaming.py

@@ -18,21 +18,26 @@
 
 import contextlib
 import re
+import signal
 import threading
 
-from gcloud.credentials import get_credentials
+from gcloud import credentials
 from google.cloud.speech.v1beta1 import cloud_speech_pb2 as cloud_speech
 from google.rpc import code_pb2
 from grpc.beta import implementations
+from grpc.framework.interfaces.face import face
 import pyaudio
+from six.moves import queue
 
 # Audio recording parameters
 RATE = 16000
-CHANNELS = 1
 CHUNK = int(RATE / 10)  # 100ms
 
-# Keep the request alive for this many seconds
-DEADLINE_SECS = 8 * 60 * 60
+# The Speech API has a streaming limit of 60 seconds of audio*, so keep the
+# connection alive for that long, plus some more to give the API time to figure
+# out the transcription.
+# * https://g.co/cloud/speech/limits#content
+DEADLINE_SECS = 60 * 3 + 5
 SPEECH_SCOPE = 'https://www.googleapis.com/auth/cloud-platform'
 
 
@@ -42,7 +47,7 @@ def make_channel(host, port):
     ssl_channel = implementations.ssl_channel_credentials(None, None, None)
 
     # Grab application default credentials from the environment
-    creds = get_credentials().create_scoped([SPEECH_SCOPE])
+    creds = credentials.get_credentials().create_scoped([SPEECH_SCOPE])
     # Add a plugin to inject the creds into the header
     auth_header = (
         'Authorization',
@@ -58,33 +63,81 @@ def make_channel(host, port):
     return implementations.secure_channel(host, port, composite_channel)
 
 
+def _audio_data_generator(buff):
+    """A generator that yields all available data in the given buffer.
+
+    Args:
+        buff - a Queue object, where each element is a chunk of data.
+    Yields:
+        A chunk of data that is the aggregate of all chunks of data in `buff`.
+        The function will block until at least one data chunk is available.
+    """
+    while True:
+        # Use a blocking get() to ensure there's at least one chunk of data
+        chunk = buff.get()
+        if not chunk:
+            # A falsey value indicates the stream is closed.
+            break
+        data = [chunk]
+
+        # Now consume whatever other data's still buffered.
+        while True:
+            try:
+                data.append(buff.get(block=False))
+            except queue.Empty:
+                break
+        yield b''.join(data)
+
+
+def _fill_buffer(audio_stream, buff, chunk):
+    """Continuously collect data from the audio stream, into the buffer."""
+    try:
+        while True:
+            buff.put(audio_stream.read(chunk))
+    except IOError:
+        # This happens when the stream is closed. Signal that we're done.
+        buff.put(None)
+
+
 # [START audio_stream]
 @contextlib.contextmanager
-def record_audio(channels, rate, chunk):
+def record_audio(rate, chunk):
     """Opens a recording stream in a context manager."""
     audio_interface = pyaudio.PyAudio()
     audio_stream = audio_interface.open(
-        format=pyaudio.paInt16, channels=channels, rate=rate,
+        format=pyaudio.paInt16,
+        # The API currently only supports 1-channel (mono) audio
+        # https://goo.gl/z757pE
+        channels=1, rate=rate,
         input=True, frames_per_buffer=chunk,
     )
 
-    yield audio_stream
+    # Create a thread-safe buffer of audio data
+    buff = queue.Queue()
+
+    # Spin up a separate thread to buffer audio data from the microphone
+    # This is necessary so that the input device's buffer doesn't overflow
+    # while the calling thread makes network requests, etc.
+    fill_buffer_thread = threading.Thread(
+        target=_fill_buffer, args=(audio_stream, buff, chunk))
+    fill_buffer_thread.start()
+
+    yield _audio_data_generator(buff)
 
     audio_stream.stop_stream()
     audio_stream.close()
+    fill_buffer_thread.join()
     audio_interface.terminate()
 # [END audio_stream]
 
 
-def request_stream(stop_audio, channels=CHANNELS, rate=RATE, chunk=CHUNK):
+def request_stream(data_stream, rate):
     """Yields `StreamingRecognizeRequest`s constructed from a recording audio
     stream.
 
     Args:
-        stop_audio: A threading.Event object stops the recording when set.
-        channels: How many audio channels to record.
+        data_stream: A generator that yields raw audio data to send.
         rate: The sampling rate in hertz.
-        chunk: Buffer audio into chunks of this size before sending to the api.
     """
     # The initial request must contain metadata about the stream, so the
     # server knows how to interpret it.
@@ -105,14 +158,9 @@ def request_stream(stop_audio, channels=CHANNELS, rate=RATE, chunk=CHUNK):
     yield cloud_speech.StreamingRecognizeRequest(
         streaming_config=streaming_config)
 
-    with record_audio(channels, rate, chunk) as audio_stream:
-        while not stop_audio.is_set():
-            data = audio_stream.read(chunk)
-            if not data:
-                raise StopIteration()
-
-            # Subsequent requests can all just have the content
-            yield cloud_speech.StreamingRecognizeRequest(audio_content=data)
+    for data in data_stream:
+        # Subsequent requests can all just have the content
+        yield cloud_speech.StreamingRecognizeRequest(audio_content=data)
 
 
 def listen_print_loop(recognize_stream):
@@ -126,25 +174,36 @@ def listen_print_loop(recognize_stream):
 
         # Exit recognition if any of the transcribed phrases could be
         # one of our keywords.
-        if any(re.search(r'\b(exit|quit)\b', alt.transcript)
+        if any(re.search(r'\b(exit|quit)\b', alt.transcript, re.I)
                for result in resp.results
                for alt in result.alternatives):
             print('Exiting..')
-            return
+            break
 
 
 def main():
-    stop_audio = threading.Event()
     with cloud_speech.beta_create_Speech_stub(
             make_channel('speech.googleapis.com', 443)) as service:
-        try:
-            listen_print_loop(
-                service.StreamingRecognize(
-                    request_stream(stop_audio), DEADLINE_SECS))
-        finally:
-            # Stop the request stream once we're done with the loop - otherwise
-            # it'll keep going in the thread that the grpc lib makes for it..
-            stop_audio.set()
+        # For streaming audio from the microphone, there are three threads.
+        # First, a thread that collects audio data as it comes in
+        with record_audio(RATE, CHUNK) as buffered_audio_data:
+            # Second, a thread that sends requests with that data
+            requests = request_stream(buffered_audio_data, RATE)
+            # Third, a thread that listens for transcription responses
+            recognize_stream = service.StreamingRecognize(
+                requests, DEADLINE_SECS)
+
+            # Exit things cleanly on interrupt
+            signal.signal(signal.SIGINT, lambda *_: recognize_stream.cancel())
+
+            # Now, put the transcription responses to use.
+            try:
+                listen_print_loop(recognize_stream)
+
+                recognize_stream.cancel()
+            except face.CancellationError:
+                # This happens because of the interrupt handler
+                pass
 
 
 if __name__ == '__main__':

speech/grpc/transcribe_streaming_test.py

@@ -11,54 +11,53 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import contextlib
-import io
 import re
 import time
 
 import transcribe_streaming
 
 
-class MockAudioStream(object):
-    def __init__(self, audio_filename, trailing_silence_secs=10):
+class MockPyAudio(object):
+    def __init__(self, audio_filename):
         self.audio_filename = audio_filename
-        self.silence = io.BytesIO('\0\0' * transcribe_streaming.RATE *
-                                  trailing_silence_secs)
 
-    def __enter__(self):
-        self.audio_file = open(self.audio_filename)
+    def __call__(self, *args):
+        return self
+
+    def open(self, *args, **kwargs):
+        self.audio_file = open(self.audio_filename, 'rb')
         return self
 
-    def __exit__(self, *args):
+    def close(self):
         self.audio_file.close()
 
-    def __call__(self, *args):
-        return self
+    def stop_stream(self):
+        pass
+
+    def terminate(self):
+        pass
 
     def read(self, num_frames):
+        if self.audio_file.closed:
+            raise IOError()
         # Approximate realtime by sleeping for the appropriate time for the
         # requested number of frames
         time.sleep(num_frames / float(transcribe_streaming.RATE))
         # audio is 16-bit samples, whereas python byte is 8-bit
         num_bytes = 2 * num_frames
-        chunk = self.audio_file.read(num_bytes) or self.silence.read(num_bytes)
+        try:
+            chunk = self.audio_file.read(num_bytes)
+        except ValueError:
+            raise IOError()
+        if not chunk:
+            raise IOError()
         return chunk
 
 
-def mock_audio_stream(filename):
-    @contextlib.contextmanager
-    def mock_audio_stream(channels, rate, chunk):
-        with open(filename, 'rb') as audio_file:
-            yield audio_file
-
-    return mock_audio_stream
-
-
 def test_main(resource, monkeypatch, capsys):
     monkeypatch.setattr(
-        transcribe_streaming, 'record_audio',
-        mock_audio_stream(resource('quit.raw')))
-    monkeypatch.setattr(transcribe_streaming, 'DEADLINE_SECS', 30)
+        transcribe_streaming.pyaudio, 'PyAudio',
+        MockPyAudio(resource('quit.raw')))
 
     transcribe_streaming.main()
     out, err = capsys.readouterr()