kaldi | official location of the Kaldi project | Speech library

Before this, I checked this, snakemake's documentation, this,and this. Maybe they actually answered this question but I just didn't understand it.

In short, I create in one rule a number of files from other files, that both conform to a wildcard format. I don't know how many of these I create, since I don't know how many I originally download.

In all of the examples I've read so far, the output is directory("the/path"), while I have a "the/path/{id}.txt. So this I guess modifies how I call the checkpoints in the function itself. And the use of expand.

The rules in question are:

download_mv

textgrid_to_ctm_txt

get_MV_IDs

merge_ctms

The order of the rules should be:

download_mv (creates {MV_ID}.TEX and .wav (though not necessarily the same amount)

textgrid_to_ctm_txt (creates from {MV_ID}.TEX matching .txt and .ctm)

get_MV_IDs (should make a list of the .ctm files)

merge_ctms (should concatenate the ctm files)

kaldi_align (from the .wav and .txt directories creates one ctm file)

analyse_align (compares ctm file from kaldi_align the the merge_ctms)

upload_print_results

I have tried with the outputs of download_mv being directories, and then trying to get the IDs but I had different errors then. Now with snakemake --dryrun I get

So I have this problem. I want to use both Flask and RabbitMQ to do a microservice capable of doing some computation-heavy task. I basically wants something like the Remote procedure call (RPC) tutorial from the documentation, but with a REST Api overhead.

So I've come with that code, so far:

server.py

In my live phone speech recognition project Python's asyncio and websockets modules are used basically to enable data exchange between client and server in asynchronous mode. The audio stream which to be recognized comes to the client from inside of a PBX channel (Asterisk PBX works for that) via a local wav file that cumulates all data from answering call until hangup event. While conversation is going on, an async producer pushes chunks of call record (each of them no larger than 16 kB) to asyncio queue, so that a consumer coroutine can write data to buffer before sending to the recognition engine server (my pick is Vosk instance with Kaldi engine designed to connect using websocket interface). Once the buffer exceeds a specific capacity (for example it may be 288 kB), the data should be flushed to recognition by send function and returned (as a transcript of the speech) by recv. The real-time recognition does matter here, therefore I need to guarantee that socket operations like recv will not halt both coroutines throughout websocket session (they should be able to keep queue-based data flow until the hangup event). Let's take a look at whole program, first of all there is a main where an event loop gets instantiated as well as a couple of tasks:

I'm working on a Kaldi project about the existing example using the Tedlium dataset. Every step works well until the clean-up stage. I have a length mismatch issue. After examing all the scripts, I found the issue is in the lattice_oracle_align.sh

reference:https://github.com/kaldi-asr/kaldi/blob/master/egs/wsj/s5/steps/cleanup/lattice_oracle_align.sh

I believe the issue is line 142.

I am pretty new to c and c++, so please try explain more specific what I should do. The program tries to read files from a directory using multithreads, and store the information in a map so that it can be used later.

I have been looking for similar posts. However, I am not able to figure out.

In https://github.com/kaldi-asr/kaldi/issues/938, it said that "If you get linker errors about undefined references to symbols that involve types in the std::__cxx11 namespace or the tag [abi:cxx11] then it probably indicates that you are trying to link together object files that were compiled with different values for the _GLIBCXX_USE_CXX11_ABI macro."

The solution for undefined reference to `pthread_cancel' (add "-pthread" flag does not work either.

My code is

I have a transcription server listening for audio on a port on a remote machine. Everything works If I stream a pre-recorded audio file and stream it to the port using netcat

I'm not able to do same using mic as input. I'm trying the following but for some reason audio is not getting streamed or I can't see and transcriptions happening or maybe I'm not sure how to get the response back in python

I've been pounding my head against a wall for three days on a Python automation pipeline that takes the binary byte array of .WAV email attachments (e.g. b'RIFFm\xc1\x00\x00WAVEfmt [...]') a phone system automatically pushes, push it through some text-to-speech API like speech_recognition or some future offline Sphinx/Kaldi implementation, and send a transcript back. Ideally, this would all be handled in memory without needing to create files on disk since that seems superfluous but I'm trying to figure out anything that Pythonically moves from the audio data I have to a transcript I can send and I don't mind a little file cleanup.

The problem I'm running into is the .WAV file attachments I manually downloaded for testing and binary data I'm working with through the email API aren't playing nice with the wave dependency, with wave.open('ipsum.wav') giving an Error: unknown format: 49 and work with the speech_recognition library ends with that wave unknown format error translating into a ValueError: Audio file could not be read as PCM WAV, AIFF/AIFF-C, or Native FLAC; check if file is corrupted or in another format.

Manually converting the local files I have into .wavs using an online file conversion tool seems to fix the issue in a way speech_recognition is willing to work with and I've managed to get a working transcript doing this (the transcript was too short for the file but that's a separate chunking issue). So the problem seems to be that wave isn't happy with how the files the phone system sends me are formatted/encoded/compressed and the solution sits somewhere in replicating how that web conversion tool encoded those test files.

I've been messing around with pydub's .export() function to try forcing it to convert to something wave likes (pydub has managed to play those files) but it seems to have taken me in a circle and I wind up back where I started with the error traceback discussed above. The ideal solution probably lies in some tool that manipulates the byte array of email attachments in memory but, again, I'm open to any Pythonic suggestions.

I might change up the text-to-speech framework I use from Google's somewhere down the line but the code for I've got so far for my basic implementation:

I am using the Kaldi speech recognition toolkit's "online 2-tcp-nnet3-decode-faster". The server receives raw audio and sends the text corresponding to this audio live. In other words, when using such a server, the idea is to start transcribing audio as soon as it is sent.

If the server is busy serving one client request, it cannot handle a second one. The second request will remain idle until the first transcription completes and the first client closes the connexion.

I would like to build a python client to communicate with the TCP server via websockets. I am able to create a socket connexion, however, I am still not able to determine whether the server is already serving another client so that I can try other servers on other ports or, create a new server instance on the fly.

I am using something like the snippet below. The call to connect succeeds even when the server is serving another client.

I am currently doing an internship as a data scientist in a Startup and I am supposed to search for and implement existing automatic speech recognition frameworks. I have an intermediate knowledge of python and feel a little overwhelmed with the task.

I have looked for solutions on Github and Kaldi which is commonly used for ASR was mentioned a lot. I am still however not able to install it on my computer (windows) since it apparently is made for use on Linux.

Other than that I haven't found too many feasible solutions for python and that's why I wanted to ask if you guys have any experience with automatic speech recognition and if you can recommend a framework for python?