Speech Note

Rating: 
5
Your rating: None Average: 5 (11 votes)

Experimental app for note taking with speech to text.

    Speech Note converts speech to text using DeepSpeech library and language models. All voice processing is entirely done locally on the device. Internet connection is only required for model download during app initial configuration. Speech Note respects your privacy and provides truly offline speech-to-text capability.

    DeepSpeech models for particular language can be downloaded directly from the app. Following models are currently configured for download:

    • Czech / cs
    • English / en
    • German / de
    • Spanish / es
    • French / fr
    • French (Common Voice) / fr
    • Italian / it
    • Italian (Mozilla Italia) / it
    • Polish / pl
    • Chinese / zh-CN

    The exact sources are listed here.

    The quality of speech recognition strongly depends on language model. In general it is not perfect but for some languages is surprisingly fine. I would be grateful for any feedback how good speech transcription is for individual models.

    Known issues:

    • Jolla Tablet: does not work at all because there is no x86 build for DeepSpeech library
    • Jolla 1: speech transcription is slow and sometimes app crashes due to low memory error
    • PinePhone: very unstable and sometimes causes crash of PulseAudio server

    Any comments, ideas, issue reports are highly appreciated.

    Source code: https://github.com/mkiol/dsnote
    Bugs, Feature requests: https://github.com/mkiol/dsnote/issues or just email: dsnote@mkiol.net

    Application versions: 
    AttachmentSizeDate
    File harbour-dsnote-1.0.1-1.aarch64.rpm1 MB29/04/2021 - 20:50
    File harbour-dsnote-1.0.1-1.armv7hl.rpm1012.08 KB29/04/2021 - 20:50
    File harbour-dsnote-1.2.0-1.armv7hl.rpm1018.13 KB18/09/2021 - 20:31
    File harbour-dsnote-1.2.0-1.aarch64.rpm1.01 MB18/09/2021 - 20:31
    File harbour-dsnote-1.3.0-1.aarch64.rpm1.01 MB01/10/2021 - 21:11
    File harbour-dsnote-1.3.0-1.armv7hl.rpm1 MB01/10/2021 - 21:11
    Changelog: 

    1.3.0

    • Czech language model and translation (many thanks to Lukáš Karas for the contribution)
    • New additional models: French (Common Voice), Italian (Mozilla Italia)

    1.2.0

    • Option to transcribe audio file
    • Minor UI fixes and improvements

    1.0.1

    • support for Jolla 1, Jolla C and PinePhone (alpha)
    • speech recognition accuracy is much improved thanks to DeepSpeech library update to version '0.10.0-alpha.3'
    • UI minor fixes

    Comments

    lispy's picture

    A big Thankyou for the Transcribe Audio File feature. Made my day!!!

    eson's picture

    How about more language models, Swedish in perticular? ;)

    mkiol's picture

    I've tried but unfortunately I didn't find any available DeepSpeech model for Swedish. If you find one I will be pleased to add it.

    eson's picture

    Well, knowing exactly nothing about the matter, I found these links on the net. Maybe you've already seen them or they are totally useless?

    https://github.com/AlexandrosFerles/Swedish-Language-Automatic-Speech-Re...

    https://github.com/se-asr/model

    https://medium.com/@klintcho/creating-an-open-speech-recognition-dataset...

     

    Thanks anyway for your good work as allways!

    defactofactotum's picture

    Now working on pinephone with sfos4.2. But the microphone disconnects after every input.

    Fuchur's picture

    It really is working very well and a very nice app.
    One thing I really would love to see is to be have a button on the keyboard or an own keyboard layout which would include it to the keyboard input.

    That would just be great :).

    lispy's picture

    Really works. I like it. My wife has to convert a huge audiofile to text but pushing the button for an hour sadly doesn't cut it for her. Can you imagine an audiofile import of sorts? Or maybe make the button sticky?

    mkiol's picture

    There are to modes (Settings->"Speech detection mode"). In "Automatic" mode, you don't have to hold the button. App will (in most cases ;-) automaticaly detect that speaking begins.

    defactofactotum's picture

    Thanks for the keyboard fix! It still doesn't work on my pinephone - it worked briefly in Italian but with very bad recognition, then stopped again. Another suggestion: would it be possible to add words to the database? I imagine this is probably a huge and complicated task....

    defactofactotum's picture

    Also does not work on pinephone. Suggestion for keyboard behaviour: at the moment it's possible to edit text in the middle of a line but after typing one letter the cursor snaps back to end of line. When an entire word is wrong this is very laborious.

    mkiol's picture

    Thank for suggestion.

    In the meantime, I've managed to fix Jolla 1, Jolla C and PinePhone issue. Moreover with alpha version of DeepSpeech accuracy of recognition is much improved. Stay tuned for next release :)

    ichthyosaurus's picture

    This looks very promising - I suggest that you ask for it to be included in the next community news :)!

    mkiol's picture

    Unfortunately app does not work on Jolla C (and most likely on other older devices). Sorry :(

    sashikknox's picture

    Cool, start testing... Too long time while download soeech model

    mkiol's picture

    Indeed, download time might be long. Model size for english is almost 1GB.

    Pelzlurch's picture

    For a first version indeed quite polished. Recognition quality is not brillant but quite OK - and very cool - offline!
    The only thing I noticed negatively is that there is no automatic line break.

    defactofactotum's picture

    Thanks for this! I haven't tried it much except for simple phrases but seems to work well. Maybe in the future you could make it easier to copy lines of text to other apps.

    oops just noticed now you can copy text on the pulley menu!
     I dictated this with speech note....

    mkiol's picture

    nice :D

    ziellos's picture

    Thanks a lot! Had no chance to really test speech recognition, but your app looks already very polished.