willow-inference-server

ed Amazon/Google anyways with HomeAssistant in combination with WIS (https://github.com/toverainc/willow-inference-server) and Willow autocorrect (https://github.com/toverainc/willow-autocorrect). Which gets me most of the things I cared about from an Echo without ever having anything leave my LAN. But it's a real PITA the setup. There is basically zero middleground in this industry between "Amazon/Google have a permanent microphone that listens to everything you do" and "You have to manually flash ESP32s and configure servers".

ed by the latest drivers and CUDA 12 on whatever OS you want. Nvidia knows what they're doing in terms of productization and the Jetson line should not be seen as some kind of secret hack/unlock for getting CUDA performance with gobs of RAM. In the case of LLMs I wouldn't be surprised at all if CPU beats it and at that point pickup 256GB of RAM or whatever for equivalent cost. In the end what do I care what people use, I'm offering the perspective and experience of someone who has actually used the Jetson line for many years and frequently struggled with all of these issues and more. [0] - https://github.com/toverainc/willow-inference-server/discuss...

s. It's the only market I know and have experience with. This is actually relevant to my comments about fundamental approach and architecture. Willow was designed for this market and their use cases which needless to say has substantially higher expectations than the more-or-less consumer/"tinkerer" market. I will NEVER monetize consumer/home users directly (again, I don't understand the market) but they get to benefit from our substantially better approach. All of that said I'd certainly be interested in partnering with someone to bring Willow to the consumer market commercially. I've spent my entire career with the "sales cycle measured in months with a big customer for a lot of money" approach, not the "tons of consumers paying $5 a month" approach. Everything you're describing is what we're currently focusing on. We feel we have the fundamentals of voice and audio pretty much nailed. Then the next question becomes "Great, I have performant and accurate speech in and out. What do I actually do with it?". Taking the speech recognition transcripts and matching them with flexible grammar, functionality and application routing, etc is the next significant challenge. We have some of our own logic internally that we will be building out but at this early stage we heavily rely on the command processing and intent recognition from our supported command endpoints. Nothing you're describing is impossible (of course) but Alexa (for example) has an eight year head start. It's going to take a while to reach feature parity with that. Our intention is to not only match Alexa functionality but also far surpass it by leveraging the entire open source ecosystem. [0] - https://github.com/toverainc/willow-inference-server/discuss...

s considered it does pretty well. I know you're early (we are too!) but in your demo I counted roughly six seconds of latency between the initial hello and response (and nearly 10 for "tell me a joke"). In terms of conversational voice this feels like an eternity. Again, no shade at all (believe me I understand more than most) but just something I thought I'd add from my decades of experience with humans and voice. This is why we have such heavy emphasis on reducing latency as much as possible. For an idea of just how much we emphasize this you can try our WebRTC demo[1] which can do end-to-end (from click stop record in browser to ASR response) in a few hundred milliseconds (with Whisper large-v2 and beam size 5 - medium/1 is a fraction of that) including internet latency (it's hosted in Chicago, FYI). Running locally with WIS and Willow we see less than 500ms from end of speech (on device VAD) to command execution completion and TTS response with platforms like Home Assistant. Granted this is with GPU so you could call it cheating but a $100 six year old Nvidia Pascal series GPU runs circles around the fastest CPUs for these tasks (STT and TTS - see benchmarks here[2]). Again, kind of cheating but my RTX 3090 at home drops this down to around 200ms - roughly half of that time is Home Assistant. It's my (somewhat controversial) personal opinion that GPUs are more-or-less a requirement (today) for Alexa/Echo competitive responsiveness. Speaking of latency, I've been noticing a trend with Willow users regarding LLMs - they are very neat, cool, and interesting (our inference server[3] supports LLamA based LLMs) but they really aren't the right tool for these kinds of tasks. They have very high memory requirements (relatively speaking), require a lot of compute, and are very slow (again, relatively speaking). They also don't natively support the kinds of API call/response you need for most voice tasks. There are efforts out there to support this with LLMs but frankly I find the overall approach very strange. It seems that LLMs have sucked a lot of oxygen out of the room and people have forgotten (or never heard of) "good old fashioned" NLU/NLP approaches. Have you considered an NLU/NLP engine like Rasa[4]? This is the approach we will be taking to implement this kind of functionality in a flexible and assistant platform/integration agnostic way. By the time you stack up VAD, STT, understanding user intent (while allowing flexible grammar), calling an API, execution, and TTS response latency starts to add up very, very quickly. As one example, for "tell me a joke" Alexa does this in a few hundred milliseconds and I guarantee they're not using an LLM for this task - you can have a couple of hundred jokes to randomly select from with pre-generated TTS responses cached (as one path). Again, this is the approach we are taking to "catch up" with Alexa for all kinds of things from jokes to creating calendar entries, etc. Of course you can still have a catch-all to hand off to LLM for "conversation" but I'm not sure users actually want this for voice. I may be misunderstanding your goals but just a few things I thought I would mention. [0] - https://github.com/toverainc/willow [1] - https://wisng.tovera.io/rtc/ [2] - https://github.com/toverainc/willow-inference-server/tree/wi... [3] - https://github.com/toverainc/willow-inference-server [4] - https://rasa.com/

070 is at least 5x faster than a Threadripper 5955WX. Our goal is to provide an open-source commercial voice assistant equivalent user experience and that is and will be fundamentally impossible for the foreseeable future on CPU. Slight tangent but there are users in the space who seem to be under the impression that they can use their Raspberry Pi for voice assistant/speech recognition. It's not even close to a fair fight but with the same implementation and settings a GTX 1070 is roughly 90x (nearly two orders of magnitude) faster[2] than a Raspberry Pi... Yes, all-in a machine with a GTX 1070 uses and order of magnitude more power (3w vs 30x) than a Raspberry Pi but the power cost in even countries with the most expensive power in the world results in a $2-$3/mo cost difference - which I feel, at least, is a reasonable trade-off considering the dramatic difference in usability (Raspberry Pi is essentially useless - waiting 10-30 seconds for a response makes pulling your phone out faster). [0] - https://github.com/toverainc/willow [1] - https://github.com/toverainc/willow-inference-server [2] - https://github.com/toverainc/willow-inference-server/tree/wi...

ed), etc. It's almost impossible to understand "why CUDA" unless you've fought these battles and spent time with "alternatives". I’ve been active in the space for roughly half a decade and when I look back to my early days I’m amazed what a beginner like me was able to do because of CUDA. I still routinely am. What you’re able to actually accomplish with a $1000 Nvidia card and a few lines with transformers and/or a Docker container is incredible. That said I am really looking forward to Apple stepping it up here - I’ve given up on AMD ever getting it together on GPGPU and Intel (with Arc) is even further behind. The space needs some real competition somewhere. [0] - https://github.com/toverainc/willow [1] - https://github.com/toverainc/willow-inference-server

-possible speech recognition via a variety of means (Willow, WebRTC, POST a file, integration with devices and client applications, etc). Results in hundreds of milliseconds or less for most intended speech tasks. See YouTube WebRTC demo[0]. TTS. Primarily provided for assistant tasks (like Willow!) and visually impaired users. LLM. Optionally pass input through a provided/configured LLM for question answering, chatbot, and assistant tasks. Currently supports LLaMA deriviates with strong preference for Vicuna (I like 13B). Built in support for quantization to int4 to conserve GPU memory. Support for a variety of transports. REST, WebRTC, Web Sockets (primarily for LLM). Performance and memory optimized. Leverages CTranslate2 for Whisper support and AutoGPTQ for LLMs. Willow support. WIS powers the Tovera hosted best-effort example server Willow users enjoy. Support for WebRTC - stream audio in real-time from browsers or WebRTC applications to optimize quality and response time. Heavily optimized for long-running sessions using WebRTC audio track management. Leave your session open for days at a time and have self-hosted ASR transcription within hundreds of milliseconds while conserving network bandwidth and CPU! Support for custom TTS voices. With relatively small audio recordings WIS can create and manage custom TTS voices. See API documentation for more information. Much like the release of Willow[1] last week this is an early release but we had a great response from HN and are looking forward to hearing what everyone thinks! [0] - https://www.youtube.com/watch?v=PxCO5eONqSQ [1] - https://github.com/toverainc/willow