Month: April 2021

Mitchell’s Status Report for April 10

This week I worked on debugging the transcript streaming and working in my group. The consumer was changed to be a fully  asynchronous model and the webpage now properly update when the transcript is fed in. There were also web socket instabilities that were debugged. For the group work, we met Mon., Wed., Fri., Sat, and will meet Sunday for multiple hours each time. I mostly helped to test the system and tweak changes like transcript autoscrolling.

From now on my schedule is the same as the group’s schedule. We think we have a demo prepared at this point, but we will perform further stress testing on the demo and mult-mic interaction.

Team Report for April 10th

Last week our team had finished our individual components, Cambrea finished audio streaming between raspberry pi and server, Ellen finished speech to text, and speaker identification ML, and Mitchell finished the database and website setup and transcript streaming.    This week our team has focused heavily on complete system integration and testing.  We have completed the connections between all components and have real transcript streaming to the website of the input audio from users.  We are currently working on improving our speaker identification during the setup phase of the meeting.

Our main risk for right now is that the direction of arrival data from the speaker can have a small fluctuation during the time that a speaker changes or during the time that a speaker is speaking.  This fluctuation is mainly affecting our setup phase for when we register speakers to be identified by the speaker identification ML.  During the setup phase we rely on the DOA to determine if a new speaker should be registered with the system.  We are currently fixing and testing this by only saying that a new speaker should be registered if a new DOA is detected and also if there is a significant amount of audio  coming from that direction.  This will work to ignore small fluctuations of DOA.

 

Updated Gantt Chart Link

https://docs.google.com/spreadsheets/d/1eeHeut41JF_Ju9Ys14n_sOLiZGYdpl4HtEale2ySasY/edit?usp=sharing

For our updated gantt chart, we have completed the task “Integration of RPi + Transcript + ML” half a week early so we are currently working on the task “Complete System Debugging / Testing”.  Our current “complete system” for the demo uses a computer as the intermediary server that handles the speaker ML and transcript streaming.  Ellen is working on some ML improvements for integration and testing at this time. After the demo Cambrea and Mitchell will start the migration to using the AWS Server instead of our local computer for the task “AWS Migration”.   For our latency testing and  standard testing we will complete these during week 12 and 13  after we have migrated to AWS.  We will start revising our final report the during week 11, and work on our final presentation and video during weeks 12-14.

 

 

Ellen’s Status Report for April 10

This week has been a busy one for our team! Over the weekend I wrote audio-accumulation for the UDP side of the webserver and accumulated raw audio instead of files. Then from Monday forward our focus was on integration. After every time we met I had a bunch of todo items to debug or just improve the server processing and transcript generator. We met Mon., Wed., Fri., and Sat. for multiple hours each time. My work has included small (but important) tweaks like an improved filenaming system, improved audio accumulation, improved error reporting, changing what does and does not count as a “speaker change” in the transcript and in the backend, and others I can’t recall at this point. Larger changes I made would be changing the way speech to text results are accepted from Google (only accepting “final” and not interim results) and adding more branches and states to the state machine of microphone setup.

From now on my schedule is the same as the group’s schedule as a whole. We think we have a demo prepared at this point but when we meet tomorrow we’ll do more testing on the demo we’re preparing as well as testing how two mics in the same meeting work together. By the weekend we’ll be done integrating and debugging our two-mic setup and we’ll have some idea of how well the transcript is matching our requirements/targets.

Mitchell’s Status Report for April 3

This week I worked on integrating the transcript streaming and fabricating a case for the respeaker. For the case, it was designed using VMWare Horizon to access the windows lab cluster and Solidworks to CAD the model. The transcript streaming was interfaced as a producer in the transcript updating phase in the meeting manager using asynchronous to synchronous communication to the channels layer. The consumer model will then append or rewrite the changed sections and push those updates to the web hook which will update the webpage.

At this point we’re sort of reaching the point of schedule ambiguity. I do not have any more individually-assigned tasks. I will be testing the transcript streaming further and try to get a full system that works locally on our computers.

Ellen’s Status Report for April 3

This week I worked intensively on the speaker identification subsystem. If you’ve already read the team status report you’ll know that we decided to ditch the speaker diarization packages we had previously identified as possible solutions – crucially, none of the packages provided audio streaming abstractions – and proceed with google cloud speaker diarization, which could be added to the current google speech to text request by simply tweaking a configuration variable. This created the opportunity to integrate the speech to text and speaker ID modules, so I had a lot of code to write and rewrite to both get the speaker ID module integrated and to add in the processing the way we desired it (including the DOA augmentation). Initial tests (using random vocal samples I found online) suggest to me that the system is going to work well and that DOA augmentation is actually going to be quite valuable – but I conducted this testing just yesterday, so the jury is still out, I suppose.

At this point we’re sort of reaching the point of schedule ambiguity. I would consider speaker identification to be finished as of yesterday, which is a good half-a-week early. I don’t have any more individually-assigned coding tasks. Now I’ll  be supporting integration through debugging and revision. As for what I know I want to accomplish this upcoming week: I want to rewrite the network-to-transcription queueing interface to accumulate multiple packets (if available) before sending the audio off to transcription. I also want to help Mitchell get transcript streaming fully working.

Team Status Report for April 3

This week we started to wrap up our individual modular code-writing and looked towards integration and demoing. As we tested and composed some of the most complex sections of the solution, we had to do some problem solving that led to design changes and improvements.

There are two notable design changes. The first is with regards to audio i/o on the raspberry pi. Our initial design involved passing audio through multiple queue data structures as processing was performed on it and it was prepared for networking. However, we found that this caused latency issues that resulted in sporadic and choppy audio. With the queueing being too slow, the audio stream now doesn’t go through any intermediary data structures before being networked, and audio processing is isolated on the webserver, as transcript preprocessing.

The other change pertains to the speech to text and speaker identification ML modules inside the transcript generator. We decided to utilize google cloud speaker identification because this would allow us to do all of the ML processing in a single integrated transaction. And this further allowed us to integrate the two modules into a single module. This also decreases the amount of multithreading.

Now we’re looking towards full integration of all components with each other. We’re doing local testing before deploying to AWS, which creates the risk that what works on our own computers doesn’t work once we deploy properly.