This tutorial shows you how to transcribe the audio from a file on your computer using Dialogflow. Dialogflow was made to be used for conversations, but as it sends a transcription of what was said we can use it to transcribe audio as well.
NOTE: This is not necessarily the best way to transcribe audio.
This demo is here for two reasons:
To provide a platform independent way to test your Dialogflow setup
To demonstrate how to work with audio and more complex Dialogflow setups in the framework
In addition to the given preliminaries, you will also need to have PyAudio installed on your virtual environment.
Approach
This tutorial will show you how to convert audio to text. We’ll split this up into a couple parts
Converting an audio file to a
.wavfileStarting the Dialogflow component
Transcribing the audio file
Converting to .wav format
To be able to read the audio in python, its easiest to convert it to a .wav file. Depending on which file type you have this might need to be done differently, but here is an example using ffmpeg. Make sure to convert it to mono 16bit PCM little-endian audio (this is what pcm_s16le means).
ffmpeg -i my_audio.mp3 -codec:a pcm_s16le -ac 1 -ar 44100 my_audio.wav
Installing and starting Dialogflow
First, start the SIC Dialogflow service, you should see something like this:
[SICComponentManager 192.168.0.181]: INFO: Manager on device 192.168.0.181 starting [SICComponentManager 192.168.0.181]: INFO: Starting component manager on ip "192.168.0.181" with components: [SICComponentManager 192.168.0.181]: INFO: - DialogflowService
Getting a key
If you don’t already have a key, check out Getting a google dialogflow key
If everything went right, you should have a your_dialogflow_key.json
Transcribing the audio
Alright! Now that we have everything set up we can start transcribing the audio.
Just to be sure, make sure you have:
The dialogflow component is running
You have a dialogflow key
A
.wavaudio file in the folder you are working in
In a new python file copy the following code:
import threading
import pyaudio
import wave
import json
from sic_framework.core.message_python2 import AudioMessage
from sic_framework.services.dialogflow.dialogflow import DialogflowConf, GetIntentRequest, Dialogflow, \
StopListeningMessage, QueryResult, RecognitionResult
To read the wave file we can use the python wave library. This will read the file as raw bytes, which is what dialogflow will expect from us.
# Read the wav file
wavefile = wave.open('office_top_short.wav', 'rb')
samplerate = wavefile.getframerate()
print("Audio file specs:")
print(" sample rate:", wavefile.getframerate())
print(" length:", wavefile.getnframes())
print(" data size in bytes:", wavefile.getsampwidth())
print(" number of chanels:", wavefile.getnchannels())
print()
Now we get to more interesting stuff. The Dialogflow component will send back a lot of information, so we will have to handle that, and extract the transcription.
First, we’ll create an event. We’ll set this event whenever Dialogflow has detected the end of a sentence. That way we can ask Dialogflow to listen to the next immediately after. Its easiest to use a threading.Event, because Dialogflow will signal the end of a sentence at an arbitrary point.
The on_dialog function handles setting this event. It also will print the partial transcript intermittently and once dialogflow has chosen a final transcript we’ll add this to the list.
# set up the callback and variables to contain the transcript results
# Dialogflow is not made for transcribing, so we'll have to work around this by "faking" a conversation
dialogflow_detected_sentence = threading.Event()
transcripts = []
def on_dialog(message):
if message.response:
t = message.response.recognition_result.transcript
print("\r Transcript:", t, end="")
if message.response.recognition_result.is_final:
transcripts.append(t)
dialogflow_detected_sentence.set()
Now we can set up dialogflow. We do this by first reading in our json key
# read you keyfile and connect to dialogflow
keyfile_json = json.load(open("path/to/your_keyfile_here.json"))
And then we can create a configuration for the dialogflow component. Make sure to set the proper sample rate!
conf = DialogflowConf(keyfile_json=keyfile_json,
sample_rate_hertz=samplerate, )
dialogflow = Dialogflow(conf=conf)
We’ll direct the output message’s produced by dialogflow to the on_dialog function by registering it as a callback.
dialogflow.register_callback(on_dialog)
To get a sense of what dialogflow is hearing, we’ll also play the sound on our own speakers.
# OPTIONAL: set up output device to play audio along transcript
p = pyaudio.PyAudio()
output = p.open(format=pyaudio.paInt16,
channels=1,
rate=samplerate,
output=True)
With everything set up, we can start to ask dialogflow to detect a sentence! We do this using dialogflow.request(GetIntentRequest(), block=False). Non blocking is important here, because we need to keep sending audio (and not wait for some result, which will not occur because no audio is sent). Every time dialogflow detects a sentence, we ask it to listen for the next one!
# To make dialogflow listen to the audio, we need to ask it to "listen for intent".
# This means it will try to determine what the intention is of what is being said by the person speaking.
# Instead of using this intent, we simply store the transcript and ask it to listen for intent again.
print("Listening for first sentence")
dialogflow.request(GetIntentRequest(), block=False)
# send the audio in chunks of one second
for i in range(wavefile.getnframes() // wavefile.getframerate()):
if dialogflow_detected_sentence.is_set():
print()
dialogflow.request(GetIntentRequest(), block=False)
dialogflow_detected_sentence.clear()
# grab one second of audio data
chunk = wavefile.readframes(samplerate)
output.write(chunk) # replace with time.sleep to not send audio too fast if not playing audio
message = AudioMessage(sample_rate=samplerate, waveform=chunk)
dialogflow.send_message(message)
When we’re done we’ll write the output to a file and clean up dialogflow.
dialogflow.send_message(StopListeningMessage())
print("\n\n")
print("Final transcript")
print(transcripts)
with open('transcript.txt', 'w') as f:
for line in transcripts:
f.write(f"{line}\n")
output.close()
p.terminate()
Thats the code! Run your file like so:
cd sic_framework/tests python3 demo_transcribe_with_dialogflow.py
The output should look something like this:
Audio file specs: sample rate: 44100 length: 4505992 data size in bytes: 2 number of chanels: 1 Component not already alive, requesting DialogflowService from manager 192.168.0.181 [DialogflowService 192.168.0.181]: INFO: Started component DialogflowService Listening for first sentence Transcript: I can't believe I started the fire Transcript: a brown Transcript: I'm taking two so I can parcel them up and eat them at my leisure later on much healthier Final transcript ["I can't believe I started the fire", ' a brown']
And the transcript should be stored in transcript.txt!
When the transcript is done, you might get some errors about google not receiving new request, like
google.api_core.exceptions.InvalidArgument: 400 Did not receive any new request for 1m.
We are still working on how to properly end a conversation after it is done, but Google’s documentation has little mention of this. If you find it, let us know!