I'm trying out this Cognitive Service speech-to-text CLI:
https://learn.microsoft.com/en-us/azure/cognitive-services/speech-service/quickstarts/speech-to-text-from-file?tabs=linux%2Cbrowser%2Cwindowsinstall
&pivots=programmer-tool-spx
I can generate an output file, but the transcript is just one very long line.
Is there a way to have some line breaks somehow? Maybe when the service detects a pause, it inserts a line break?
Also when I try to use an MP3 file (using --format mp3), I get this error: ERROR: Exception with an error code: 0x29 (SPXERR_GSTREAMER_NOT_FOUND_ERROR) Do I need to install anything else for MP3 to work?
Thank you very much.
Hi, thanks for the updates. In response to your first question, we current don't have paragraph level support. However, you can easily separate the output text based on sentence ending with
period
and
timestamp
to decide the pause between two sentences before breaking them into paragraphs. I have provided this feedback internally to the product team but feel free to share this request on
uservoice
forum so you and others can up-vote to enable product team prioritize this feature. Thanks.
For anyone who stumbles upon this and have the same question about the error I posted above, there actually is a documentation for it:
https://learn.microsoft.com/en-us/azure/cognitive-services/speech-service/how-to-use-codec-compressed-audio-input-streams?tabs=debian&pivots=programming-language-csharp
So after further tinkering I found out you have to use this version of GStreamer, otherwise the necessary library file (libgstreamer-1.0-0.dll) will not be available:
https://gstreamer.freedesktop.org/data/pkg/windows/1.15.1/
I used this file:
gstreamer-1.0-x86_64-1.15.1.msi
Using the Speech CLI (spx), you can output different types of files, that will have the text "chunked" in different ways...
By default, spx will output "all" recognized events, concatenated, in a file like "output.*.tsv"
Using
--output each
command line options, it will also produce a file with "each" recognized event, thus, the text will be in chunks
Uy using
--output batch json
command line option, it will also/or instead produce a file with "all" recognized text concatenated
See
spx help recognize output
for more details on each of the three options.
Thus, for your specific question, you can do this:
spx recognize --file example.wav --output each recognized text --output each file output.tsv --output each tsv file has header false
