VibeVoice: Open-source frontier voice AI

1 day ago (github.com)

This is not a new model. Also, it hallucinates a lot. Also, it's very heavy and slow in inference. It's also bad in multilingual.

Edit: I'm talking purely about speech to text (STT). Not sure about the other things this can do.

  • It has some perks, is a bit more expressive in some cases, but overall is trained on really noisy data, uses more memory, and isn't that fast - I'm talking about the (7b?) version that they released then removed quickly (vibevoice-community on github) - I still use chatterbox turbo and sometimes qwen TTS.

  • Yeah, I don't get why it is suddenly getting so much attention today, it is all over twitter too

  • It is not good for text to speech (TTS) as well. I am trying it for few days. First of all 1.5B model documentation is not there. 0.5B realtime is shit model. I was converting text, line by line and it was randomly adding music and couldn't handle special characters like "…".

    I really disappointed with this model to say the least.

    • The 7B parameter Vibevoice TTS model is still the most impressive local TTS model i've tried. It was pulled by Microsoft a few days after its release due to "abuse potential" but it can be found in various community maintained huggingface repos.

    • yep, it seems this was trained on large amount of podcasts with ad jingles or phone call queues with elevator music. I was also pretty disappointed to run the TTS last week.

I think we should stop calling this type of models open source. They are indeed "open weight." The training code is proprietary and never revealed.

https://github.com/microsoft/VibeVoice/issues/102

  • I'm reserving that complaint for "open source" models which are released under non-open-source licenses.

    I care that I know what I can DO with the project when I see it described as "open source".

    • > I care that I know what I can DO with the project when I see it described as "open source".

      Yes, the first of which is that you should be able to build it from source. Which requires the source code, and in this case data.

      6 replies →

    • That would be “permissive license”

      Maybe we should have a little cue card for models: vendor/name, size, open weights, open source, permissive license.

      It’s simple enough an idea.

  • > we should stop calling this type of model open source. They are indeed "open weight”

    This ship has sailed. It’s now in the same category as hacker/cracker and the pronunciation of GIF.

  • Devils advocate here: I can give you a binary of my open source MIT code and never phone you the code. The code is still MIT licensed, and open source. You just have no access to it.

    That said, I entirely agree that MS is misrepresenting their openness here, which isn’t in the least surprising.

    • ? Do you know what “source” means in open source? Like, what is the source of the binary? It’s the code. That’s the source in open source.

      4 replies →

    • In their defense, most everyone else does the same thing. They still shouldn't do it, but at least they're not the trendsetter here (though they are contributing to the ongoing problem)

  • At least it's MIT licensed! As much as non-open training data irks me, restrictive licensing irks me more!

  • What you said makes a lot of sense. Free software should not be confused with open source

  • I mean, you have "AI" which means just about anything in marketing speak, "Agentic" is kind of becoming similar, hopefully they don't goof that one too badly, would be nice to know what you are trying to sell me. Used to be "Cloud" meant storage not just hosting (I guess it still does).

    Then there's "Smart" in front of Car, Phone, TV, and so on... Meaning different things.

    I do think "Open Weight" should be more commonly used. There's definitely communities that spring up that build the training infrastructure and inference infrastructure around open models on the other hand.

  • Open weights is not exactly right either because we do get source of the software that uses those open weights.

    Maybe open inference?

    But we often also get source code for fine tunning the model.

    So maybe it's closer to open source than to anything else?

    Isn't it a bit like not calling a game open source because engine tooling used to made it isn't open source and they didn't publish .psd files with asset designs?

  • I'm genuinely torn on this one; I get technically why not, but why I think I have no problem with it is the wishy-washiness of "open source" generally.

    As I teach this stuff to people newer to this tech, it's probably just easier and more helpful to refer to the wide array of "stuff you can just download and use yourself" as "open-source" and then after that, go deeper and talk about why Stallman was right, how "Free Software" was first. etc.

Isn't this project the one Microsoft published but then soon after pulled it for security/safety reasons? What has changed since then?

  • Look at the "News" section in the readme - The original TTS model is gone from this repo (you can still find it other places), but the SST/ASR, long form TTS, and streaming TTS models are newer.

  • It’s confusing (at least for me) because the project covers a number of things including what you are mentioning.

    • [off topic]

      When explanations get posted directly in HN comments, I imagine someone somewhere in the world is able to learn in spite of their Internet restrictions/firewalls

      People will also post their own interpretations in response to comments, and quickly find out they missed something.

      … But if you try to automate it, like include a summary under every HN post, you encourage laziness too much and are pre-chewing too heavily. Some balance here.

      [on topic]

      (OK I’m done making excuses, time to read the article… thanks for the encouragement!)

      I thought this was not explained in the readme directly but in fact I missed it. I wasn’t going to read Microsoft entire changelog! But it was substantive, thanks to sibling commenter:

      “2025-09-05: VibeVoice is an open-source research framework intended to advance collaboration in the speech synthesis community. After release, we discovered instances where the tool was used in ways inconsistent with the stated intent. Since responsible use of AI is one of Microsoft’s guiding principles, we have removed the VibeVoice-TTS code from this repository.”

Interesting to see "vibe" enshrined by the likes of Microsoft as an AI product word.

  • Especially when "vibe coded" can have a negative connotation meaning quickly put together without understanding.

    • In my mind, Vibe-anything means "some slop carelessly thrown together to ship as fast as possible." Wild that it's being used in a serious product name!

    • I’m just surprised they put the name of the e-waste slop company in their product

  • Maybe they were trying to make a pun on "Via Voice", the cursed IBM STT from the 90s?

I built speech-swift, which focuses on on-device speech processing like VibeVoice, but specifically leverages Apple Silicon's capabilities for ASR, TTS, and VAD without cloud dependency. Our ASR supports 52 languages with a real-time factor of 0.06. https://soniqo.audio/benchmarks

You have selected Microsoft Sam as the computer's default voice.

  • My friends and I had fun in the computer lab with Microsoft Sam, inputting long strings of characters to create funny sound effects. Sususususususu.

So we've really just settled on Vibe as the verb for AI then?

  • I'd be willing to bet it will be "Word of the Year" for 2026. Merriam-Webster had 'slop' for 2025, and 'polarization' for 2024. Is there a prediction market for this?

    • it'll probably be something we're not even talking about yet - we still have 7 months in which to make the world even worse

Still waiting for the open weights model that conclusively beats the multi-year old Whisper in accuracy, features, and performance.

  • It's crazy that a lot is happening in open models for stt, but there's very little progress when it comes to results, esp multilingual.

I've been using VibeVoice's ASR (speech to text) model quite intensively for the past month and have found it to be a lot more reliable and out-of-the box functional then Whisper, parakeet and other models. The fact that is has diarization built into to the model is a huge win in my book. Without that you have to run a different model just for that which adds significantly to the overall processing time vs VibeVoice which gives you reliably great results. Big fan.

The 60-minute single-pass transcription is the part that actually matters. Most ASR models chunk audio and you lose speaker continuity across boundaries. If the diarization actually holds up on hour-long recordings without drifting, thats a real solve for podcast and meeting transcription workflows.

I took a look into local options for ASR and diarization some months ago, I missed that VibeVoice now has this feature.

My conclusions back then (which only came from a shallow research on the topic and 0 real experience mind you) was that Whisper + Pyannote was the "stable" approach.

Have the VibeVoice, Voxtral, Qwen or the Nemo solutions caught up in segmentation and speaker recognition?

  • It highly depends on the sort of data you’re processing (phone calls, podcasts, meetings of more people recorded using single channel?). For NVIDIA/NeMo, check out their softformer diarization models (also streaming).

the built-in diarization is the one thing that actually caught my attention here. running whisper + pyannote separately is a pain for long recordings and the speaker continuity breaks at chunk boundaries. if this handles it in a single pass that's a real workflow improvement, regardless of how the raw accuracy benchmarks compare

I the past month or so, I added 2 models to my app Whisper Memos (https://whispermemos.com):

- Cohere Transcribe (self hosted)

- Grok Speech To Text (they provide an API, only $0.10/hr!)

They are both excellent. I'm not sure about this one. Would you like to see it in a consumer speech to text app?

  • I've had good experiences with the Mistral Voxtral models (I've used the API, but some of the model-variants are open weight)

  • Does Cohere work with longer transcripts? Do you have to do some magic to merge recordings over 35 seconds long?

  • Any non-Musk alternatives that are comparable in quality and cost?

    • Voxtral competes on price ($0.003/min) and quality. Speechmatics has best in class accuracy but is a bit more expensive ($0.004/min)

    • Our default is still OpenAI Whisper. Grok is just a choice for users who might prefer it.

What’s the current state of the art, for each of training locally and in the cloud, for learning my voice?

Microsoft has historically made poor choices in product naming, but this has to be a new low.

Microsoft continues to be completely incapable of coming up with good names for their products and services

When mixing languages, why does the English have Chinese accent and Chinese have English accent? Is it a feature or bug?

Explains most of the shit they have pushing with Windows 11. Perhaps all that bloatware was VibeVoiced too.

It would have been better if they provided not just weights, but also some frontend where it is usable as is.

Seems quite heavy for a STT model, Parakeet and Whisper are much smaller and perform great for quick dictation and transcription of longer files. I guess that's due to additional accuracy and speaker diarisation?

The TTS example clip in the repo of 'spontaneous singing' is creepy as fuck