Show HN: Open-source, native audio turn detection model

github.com

124 points by kwindla 3 days ago

Our goal with this project is to build a completely open source, state of the art turn detection model that can be used in any voice AI application.

I've been experimenting with LLM voice conversations since GPT-4 was first released. (There's a previous front page Show HN about Pipecat, the open source voice AI orchestration framework I work on. [1])

It's been almost two years, and for most of that time, I've been expecting that someone would "solve" turn detection. We all built initial, pretty good 80/20 versions of turn detection on top of VAD (voice activity detection) models. And then, as an ecosystem, we kind of got stuck.

A few production applications have recently started using Gemini 2.0 Flash to do context aware turn detection. [2] But because latency is ~500ms, that's a more complicated approach than using a specialized model. The team at LiveKit released an open weights model that does text-based turn detection. [3] I was really excited to see that, but I'm not super-optimistic that a text-input model will ever be good enough for this task. (A good rule of thumb in deep learning is that you should bet on end-to-end.)

So ... I spent Christmas break training several little proof of concept models, and experimenting with generating synthetic audio data. So, so, so much fun. The results were promising enough that I nerd-sniped a few friends and we started working in earnest on this.

The model now performs really well on a subset of turn detection tasks. Too well, really. We're overfitting on a not-terribly-broad initial data set of about 8,000 samples. Getting to this point was the initial bar we set for doing a public release and seeing if other people want to get involved in the project.

There are lots of ways to contribute. [4]

Medium-term goals for the project are:

  - Support for a wide range of languages
  - Inference time of <50ms on GPU and <500ms on CPU
  - Much wider range of speech nuances captured in training data
  - A completely synthetic training data pipeline. (Maybe?)
  - Text conditioning of the model, to support "modes" like credit card, telephone number, and address entry.
If you're interested in voice AI or in audio model ML engineering, please try the model out and see what you think. I'd love to hear your thoughts and ideas.

[1] https://news.ycombinator.com/item?id=40345696

[2] https://x.com/kwindla/status/1870974144831275410

[3] https://blog.livekit.io/using-a-transformer-to-improve-end-o...

[4] https://github.com/pipecat-ai/smart-turn#things-to-do

pzo 2 days ago

I will have a look at this. Played with pipecat before and it's great, switched to sherpa-onnx though since I need something that compile to native and can run on edge devices.

I'm not sure if turn detection can be really solved except dedicated push to talk button like in walkie-talkie. I often tried google translator app and the problem is in many times when you speaking longer sentence you will stop or slow down a little to gather thought before continuing talking (especially if you are not native speaker). For this reason I avoid converation mode in such cases like google translator and when using perplexity app I prefer the push to talk button mode instead of new one.

I think this could be solved but we would need not only low latency turn detection but also low latency speech interruption detection and also very fast low latency llm on device. And in case we have interruption good recovery that system know we continue last sentence instead of discarding previous audio and starting new etc.

Lots of things can be improved also regarding i/o latency, like using low latency audio api, very short audio buffer, dedicated audio category and mode (in iOS), using wired headsets instead of buildin speaker, turning off system processing like in iphone audio boosting or polar pattern. And streaming mode for all STT, transport (using using remote LLM), TTS. Not sure if we can have TTS in streaming mode. I think most of the time they split by sentence.

I think push to talk is a good solution if well designed: big button in place easily reached with your thumb, integration with iphone action button, using haptic for feedback, using apple watch as big push button, etc.

  • genewitch a day ago

    Whisper can chunk on word boundaries or split on word boundaries. The speaker diarization stuff, I can't remember the name offhand, but it also can split on the word boundaries since it needs to identify speakers per words.

foundzen 3 days ago

I got most of my answers from the README. Well written. I read most of it. Can you share what kind of resources (and how much of them) were required to fine tune Wav2Vec2-BERT?

  • kwindla 3 days ago

    It takes about 45 minutes to do the current training run on an L4 GPU with these settings:

        # Training parameters
    "learning_rate": 5e-5,
    "num_epochs": 10,
    "train_batch_size": 12,
    "eval_batch_size": 32,
    "warmup_ratio": 0.2,
    "weight_decay": 0.05,

    # Evaluation parameters
    "eval_steps": 50,
    "save_steps": 50,
    "logging_steps": 5,

    # Model architecture parameters
    "num_frozen_layers": 20
    I haven't seen a run do all 10 epochs, recently. There's usually an early stop after about 4 epochs.

    The current data set size is ~8,000 samples.

remram 3 days ago

Ok what's turn detection?

  • kwindla 3 days ago

    Turn detection is deciding when a person has finished talking and expects the other party in a conversation to respond. In this case, the other party in the conversation is an LLM!

    • remram 3 days ago

      Oh I see. Not like segmenting a conversation where people speak in turn. Thanks.

      • password4321 2 days ago

        Speaker diarization is also still a tough problem for free models.

      • whiddershins 2 days ago

        huh. how is analyzing conversations in the manner you described NOT the way to train such a model?

        • remram 2 days ago

          Did you reply to the wrong comment? No one is taking about training here.

  • ry167 3 days ago

    Detecting when one user of a conversation has finished talking.

    It’s a big deal for detecting human speech when interacting with LLM systems

  • woodson 3 days ago

    It’s often called endpoint detection (in ASR).

    • lelag 2 days ago

      Yes, weird that they didn't use that term for this project.

      • kwindla 2 days ago

        I've talked about this a lot with friends.

        Endpoint detection (and phrase endpointing, and end of utterance) are terms from the academic literature about this, and related, problems.

        Very few people who are doing "AI Engineering" or even "Machine Learning" today know these terms. In the past, I argued that we should use the existing academic language rather than invent new terms.

        But then OpenAI released the Realtime API and called this "turn detection" in their docs. And that was that. It no longer made sense to use any other verbiage.

        • mncharity 2 days ago

          Re SEO, I note "utterance" only occurs once, in a perhaps-ephemeral "Things to do" description.

          To help with "what is?" and SEO, perhaps something like "Turn detection (aka [...], end of utterance)"... ?

        • lelag 2 days ago

          Thank for the explanation. I guess it makes some sense, considering many people with no nlp background are using those models now…

xp84 3 days ago

I'm excited to see this particular technology developing more. From the absolute worst speech systems such as Siri, who will happily interrupt to respond with nonsense at the slightest half-pause, to even ChatGPT voice mode which at least tries, we haven't yet successfully got computers to do a good job of this - and I feel it may be the biggest obstacle in making 'agents' that are competent at completing simple but useful tasks. There are so many situations where humans "just know" when someone hasn't yet completed a thought, but "AI" still struggles, and those errors can just destroy the efficiency of a conversation or worse, lead to severe errors in function.

zamalek 3 days ago

As an [diagnosed] HF autistic person, this is unironically something I would go for in an earpiece. How many parameters is the model?

written-beyond 3 days ago

Having reviewed a few turn based models your implementation is pretty inline with them. Excited to see how this matures!

  • kwindla 3 days ago

    Can you say more? There's not much open source work in this domain, that I've been able to find.

    I'm particularly interested in architecture variations, approaches to the classification head design and loss function, etc.

lostmsu 3 days ago

Does this support multiple speakers?

  • kwindla 2 days ago

    In general, for realtime voice AI you don't want this model to support multiple speakers because you have a separate voice input stream for each participant in a session.

    We're not doing "speaker diarization" from a single audio track, here. We're streaming the input from each participant.

    If there are multiple participants in a session, we still process each stream separately either as it comes in from that user's microphone (locally) or as it arrives over the network (server-side).

fdsd 3 days ago

[dead]