Awesome
TaSTT: A deliciously free STT
TaSTT (pronounced "tasty") is a free speech-to-text tool for VRChat. It uses a GPU-based transcription algorithm to turn your voice into text, then sends it into VRChat via OSC.
To get started, download the latest .zip from the releases page.
Contents:
Made with love by yum_food.
Usage and setup
Download the latest .zip from the releases page.
Please join the discord to share feedback and get technical help.
To build your own package from source, see GUI/README.md.
Basic controls:
- Short click to toggle transcription.
- Medium click to hide the text box.
- Hold to update text box without unlocking from worldspace.
- Medium click + hold to type using STT.
- Scale up/down in the radial menu.
Design philosophy
- All language services are performed on the client. No network hops in the critical path.
- Priorities (descending order): reliability, latency, accuracy, performance, aesthetics.
- No telemetry of any kind in the app. github and discord are the only means I have to estimate usage and triage bugs.
- Permissive licensing. Users should be legally entitled to hack, extend, relicense, and profit from this codebase.
Features
- Works with the built-in chatbox (usable with public avatars!)
- Customizable board resolution, up to ridiculous sizes.
- Lighweight design:
- Works with VRC native chatbox - works with any avatar without modification
- Custom textbox requires as few as 65 parameter bits
- Transcription doesn't destroy your frames in game since VRChat is heavily CPU bound. Performance impact when not speaking is negligible.
- Performant: uses CTranslate2 inference engine with GPU support and flash-attention
- Browser source. Use with OBS!
- Multi-language support.
- Whisper natively supports transcription in 100 languages.
- A local translation algorithm (Meta's NLLB) enables translating into 200 other languages with good-ish accuracy (BLEU scores typically around 20-35) and low latency.
- Customizable:
- Control button may be set to left/right a/b/joystick.
- Text filters: lowercase, uppercase, uwu, remove trailing period, profanity censoring.
- Many optional quality-of-life features:
- Audio feedback: hear distinct beeps when transcription starts and stops.
- May also enable in-game noise indicator, to grab others' attention.
- Custom chatbox features:
- Free modular avatar prefab available here.
- Resizable with a blendtree in your radial menu.
- Locks to world space either when summoned (default) or when done speaking.
- Unicode variant (supporting e.g. Chinese and Japanese) is available through the app's Unity panel.
- Privacy-respecting: transcription is done on your GPU, not in the cloud.
- Hackable.
- From-scratch implementation.
- Free as in beer.
- Free as in freedom.
- MIT license.
Bad parts
I think that any ethical software project should disclose what sucks about it. Here's what sucks about this project:
- The app UI looks like trash. Only you will see it, so I don't think this really matters. (Electron rewrite when?)
- The app is HUGE. This mostly stems from the bundled NVIDIA CUDNN .dll's
(~1.0GB) and portable git (~500 MB).
- NVIDIA's DLLs should be statically linked into ctranslate2. That probably means doing our own build of ctranslate2... yuck.
- Portable git can probably be stripped down. It includes a full mingw environment responsible for the vast majority of the size, which we almost certainly don't need.
- The app doesn't start automatically with steamvr (TODO do this)
- The app starts in a weird state where it's transcribing and doesn't really back off correctly. Press the controller keybind once to stop transcription then again to put it into a normal state.
- The backend Unity code is pretty gory. (This is largely irrelevant to end users, since end users mostly use the VRC-native chatbox or the modular avatar prefab.) I have a burning disdain for C# so I wrote a scuffed "animator as code" library (libunity.py) in Python. This includes a lot of crazy shit like a multiprocess YAML parser and a ton of macro-like string manipulation/concatenation. We should just use the upstream C# animator as code library.
- The app doesn't include any version numbers, so debugging version-specific issues can be tough (TODO fix this)
Requirements
System requirements:
- ~2GB disk space
- NVIDIA GPU with at least 2GB of spare VRAM.
- You can run it in CPU mode, but it's really slow and lags you a lot more, so I wouldn't recommend it.
- I've tested on a 1080 Ti and a 3090 and saw comparable latency.
- SteamVR.
Avatar resources used by custom chatbox:
- Tris: 12
- Material slots: 1
- Texture memory: 340 KB (English), 130 MB (international)
- Parameter bits: 65-217 (configurable; more bits == faster paging)
- Menu slots: 1
Motivation
Many VRChat players choose not to use their mics, but as a practical matter, occasionally have to communicate. I want this to be as simple, efficient, and reliable as possible.
There are existing tools which help here, but they are all imperfect for one reason or another:
- RabidCrab's STT costs money and relies on cloud-based transcription. Because of the reliance on cloud-based transcription services, it's typically slower and less reliable than local transcription. However, the accuracy and speed of cloud AI models has improved radically since late 2022, so this is probably the best option if money and privacy don't matter to you.
- The in-game text box is not visible in streamer mode, and limits you to one update every ~2 seconds, making it a poor choice for latency-sensitive communication.
- KillFrenzy's AvatarText only supports text-to-text. It's an excellent product with high-quality source code, but it lacks integration with a client-side STT engine.
- I5UCC's VRCTextboxSTT makes KillFrenzy's AvatarText and Whisper kiss. It's the closest spiritual cousin to this repository. The author has made incredible sustained progress on the problem. Definitely take a look!
- VRCWizard's TTS-Voice-Wizard also uses Whisper, but they rely on the C# interface to Const-Me's CUDA-enabled Whisper implementation. This implementation does not support beam search decoding and waits for pauses to segment your voice. Thus it's less accurate and higher latency than this project's transcription engine. It supports more features (like cloud-based TTS), so you might want to check it out.
Why should you pick this project over the alternatives? This project is mature, low-latency (typically 500-1000 ms end-to-end in game under load), reliable, and accurate. There is no network hop to worry about and no subscription to manage. Just download and go.
Design overview
These are the important bits:
TaSTT_template.shader
. A simple unlit shader template. Contains the business logic for the shader that shows text in game.generate_shader.py
. Adds parameters and an accessor function to the shader template.libunity.py
. Contains the logic required to generate and manipulate Unity YAML files. Works well enough on YAMLs up to ~40k documents, 1M lines.libtastt.py
. Contains the logic to generate TaSTT-specific Unity files, namely the animations and the animator.osc_ctrl.py
. Sends OSC messages to VRChat, which it dutifully passes along to the generated FX layer.transcribe_v2.py
. Uses OpenAI's whisper neural network to transcribe audio and sends it to the board using osc_ctrl.
Parameters & board indexing
I divide the board into several regions and use a single int parameter,
TaSTT_Select
, to select the active region. For each byte of data
in the active region, I use a float parameter to blend between two
animations: one with value 0, and one with value 255.
To support wide character sets, I support 2 bytes per character. This can be configured down to 1 byte per character to save parameter bits.
FX controller design
The FX controller (AKA animator) is pretty simple. There is one layer for each sync parameter (i.e. each character byte). The layer has to work out which region it's in, then write a byte to the correct shader parameter.
From top down, we first check if updating the board is enabled. If no, we stay in the first state. Then we check which region we're in. Finally, we drive a shader parameter to one of 256 possible values using a blendtree.
The blendtree trick lets us represent wide character sets efficiently. The number of animations required increases logarithmically with the size of the character set:
(N bytes per character) = ceil(log2(size of character set))
(total animations) =
(2 animations per byte) *
(N bytes per character) *
(M characters per region)
Contributing
Contributions welcome. Send a pull request to this repository.
See GUI/README.md for instructions on building the GUI.
Ping the discord if you need help getting set up.
Roadmap
Milestone 1: STT Personally usable
Status: COMPLETE.
Scope: The speech-to-text may be used by one developer intimately familiar with its inner workings. Environment is not encapsulated.
Completed at commit 8326dee0bf01956.
Milestone 2: STT Generally usable
Status: COMPLETE.
Scope: The speech-to-text is used by at least one user not familiar with its inner workings. Dependency management is mostly handled mechanically. The app can be controlled using a GUI.
Completed at commit 1f15133dd985442, AKA release 0.10.0.
Milestone 3: STT Generally performant
Status: COMPLETE.
Scope: The speech-to-text may be used on resource constrained systems.
I'm looking at Const-Me/Whisper as the transcription backend. I have measured terrible accuracy when using the VAD-segmented transcription path vs. using the file-based non-VAD-segmented transcription path (~15x higher edit distance on the same recording of the Bill of Rights). Beam search has not measurably improved the file-based transcription path. It remains to be seen if VAD segmentation is the failure source, or if it's caused by the inference layer being unable to "second guess" itself (previous transcriptions cannot be edited in the current architecture), or something else.
Completed at commit 1f2e5c6cf16e7e7, AKA release 0.11.2.
Milestone 4: Enable non-VRChat use cases
Status: COMPLETE.
Scope: The speech-to-text may be used as a tool for usecases outside of VRChat.
Streamers could use the STT as an OBS browser source. VR players could use it to type into arbitrary text fields (voice-driven keyboard device). MMO players could also use the voice-driven keyboard (speak -> preview -> rapid commit?) while raiding.
Completed at commit 7a576bcac1c37c3, AKA release 0.13.1.
Milestone 5: Integration into other tools
Status: NOT STARTED.
Scope: Integrate performant client-side transcription into other STT tools.
Once performant client-side transcription is implemented, there is no reason to keep it locked away inside one project. Other projects making different tradeoffs (such as relying on cloud services for TTS) could benefit from this functionality, driving down costs and latency for users. In particular, I think that there is value in integrating with TTS-Voice-Wizard.
TaSTT is about providing performant, commoditized, user-owned STT services. I have no interest in using cloud services to provide any functionality. Instead of extending this project to do that, the best way to spread the love is to partner with (contribute to) projects that do.
Completion
This project will probably reach a stable state and then go into maintenance. The efforts described above are the major milestones I plan to implement. Small features and bugfixes will likely continue in the "completed" state.
Backlog
- Better Unity integrations
- Port all scripts to Unity-native C# scripts.
Support appending to existing FX layers.DONE- Use VRCSDK to generate FX layer instead of generating the serialized files.
- In-game usability features.
Resizing (talk to friends far away).DONEBasic toggles (hide it when not needed).DONEWorld mounting (leave it in a fixed position in world space).DONEAvatar mounting (attach it to your hand)DONE.Controller triggers (avoid having to use the radial menu every time you want to speak).DONE
- General usability features.
Error detection & correction.DONEText-to-text interface. Type in terminal, show in game.DONESpeech-to-text interface. Speak out loud, show in game.DONETranslation into non-English. Whisper natively supports translating N languages into English, but not the other way around.DONE- Display text in overlay. Enables (1) lower latency view of TaSTT's transcription state; (2) checking transcriptions ahead of time; (3) checking transcriptions without having to see the board in game.
- TTS. Multiple people have requested this. See if there are open source algorithms available; or, figure out how to integrate with
Save UI input fields to config file. Persist across process exit. It's annoying having to re-enter the config every time I use the STT.DONECustomizable controller bindings. Someone mentioned they use left click to unmute. Let's work around users, not make them change their existing keybinds.DONE- One-click upgrade. Fetch latest stable release, copy over virtual env and UI configs, relaunch.
Browser source for OBS. Blocker: the transcription layer doesn't handle long pauses well.DONE- Test suite. Some long representative transcripts with mechanical word error rate (WER) calculation.
- Optimization
Utilize the avatar 3.0 SDK's ability to drive parameters to reduce the total # of parameters (and therefore OSC messages & sync events). Note that the parameter memory usage may not decrease.DONEOptimize FX layer. We have 14k animations and a 1.2 million line FX layer. Something must be rethought to bring these numbers down.DONEImplement multicore YAML parsing. This will make working with large animators much more practical.DONETranscription engine sleep interval increases exponentially up to 1-2 seconds, then jumps back to a short interval once speech is detected. This should significantly cut down on idle resource consumption. Perhaps there's even a more efficient way to detect the odds that anything is being said, which we could use to gate transcription.DONE- There are ~64k words in the English language. We could encode each word using a 16-bit int. On the other hand, suppose you represented each character using 7 bits per character and transmitted words character-by-character. The average word length is 4.7 characters, and we send ~1 space character per word. Thus the expected bits per word in an optimized version of today's encoding scheme is (5.7 * 7) == 39.9 bits. The other encoding scheme is thus ~2.5 times more efficient. This could be used to significantly speed up sync times. (Thanks, Noppers for the idea!)
Use Const-Me/Whisper for transcription.WON'T DOImplement beam search in Const-Me/Whisper.WON'T DO
- Bugfixes
The whisper STT says "Thank you." when there's no audio?DONE- JP and CN transcription does not work in the GUI due to encoding issues.
- Shine
- Smooth scrolling.
Infinite scrolling.DONESound indicator, maybe like animal crossing :)DONESupport texture-based PBR shadingDONE