Speech-to-Text | Flask application that takes speech

I have an app that involves an Alexa like digital assistant.

We are successfully receiving input for our Speech-To-Text engine with AVAudioEngine, and then using NLP we are interpreting that text into requests and responding.

Once we receive a command, we acknowledge that command with audio, played through AVAudioPlayer.

This wasn't a problem on our simulators, but on device we have noticed that if we play a sound any time after the AVAudioEngine has been initialized, the audio does not play on the device. The functions do not fail, and our audioPlayerDidFinishPlaying fires at the appropriate time, but no audio is heard.

If we play audio when the app first launches without initializing the recording, the audio plays fine. If we stop the recording and wait one to many seconds later to play an audio file, no audio is ever heard.

We need to be able to stop the input when a command is recognized, play an audio file in response, and then resume recording.

What are we doing wrong?

AVAudioEngine instantiation and de-initialization:

Instantiation:

I'm making a speech-to-text tool. I'm capturing audio in real time (using Web audio api from Chrome) and sending it to a server to convert the audio to text.

I'd like to extract pieces of the whole audio cause I only want to send sentences, avoiding silences. (cause the api I use has a cost). The problem is that I don't know how to convert the whole audio into pieces.

I was using MediaRecorder to capture the audio

I am having an issue with transcribing (Speech-To-Text) an audio file hosted on Azure Storage container from the Cognitive Services API.

The services are of the same resource (and I created a VNet and they are part of the same subnet).

After I take the response from there the contentUrl:

The error I get is:

We are working on a speech-to-text project. We are quite new in this field and would be very grateful if you could help us.

Our goal is to use MFCC to extract features from the audio dataset, use a CNN model to estimate the likelihood of each feature, and then use an HMM model to convert the audio data to text. All of these steps are clear to us except for the labeling. When we preprocessed the data, we divided the audio data into smaller time frames, with each frame about 45ms long and a 10ms gap between each frame.

I am going to use TIMIT dataset. I am completely confused about the labeling of the data set. I checked the TIMIT dataset and I found out the label file have 3 columns. The First one is BEGIN_SAMPLE :== The beginning integer sample number for the segment, the second one is the ending integer sample number for the segment and the last one is PHONETIC_LABEL :== Single phonetic transcription. How we used this labeling? Are the first and second columns important? Thanks for your time

I am using the Google Speech-to-Text client library for Python to convert speech using speech adaptation. I want to be able to boost phrases that fit a certain pattern. I have used this documentation to create custom classes and phrase sets and put them together into a SpeechAdaptation object.

I am using IBM's speech-to-text service and it says active. How do I deactivate it so I don't use all of my minutes? I have looked everywhere and cant find a way to deactivate it.

I am creating a speech-to-text generator gui in pyqt5. The program takes input from microphone and sets the text of the text area accordingly but everytime the user takes an input, the whole text of the text edit changes. Is there any method as how to generate the second input in a new line in the text edit. Below is the code I am using.

This is the coding part for the text edit portion

When using speech to text in Azure with dictation mode ON it recognizes words like "question mark" and returns "?". We found other words like this and were looking for complete list but were not able to find it in the documentation (https://docs.microsoft.com/en-us/azure/cognitive-services/speech-service/index-speech-to-text)

This could pertain to other speech-to-text solutions, but we happen to be using Twilio.

Is there some easy means to do matching for a numerical date from spoken user input? For example, 08/11/2020 could be spoken as 'August Eleventh Twenty Twenty' 'Zero Eight one one Twenty Twenty' or various combinations. Because of how our app works, we need an exact match.

It seems that this would be a common issue, and I am wondering if there is already a solution. Any help would be appreciated.