nshm

Yes, and it is very good one. Leading position on private leaderboard on HF: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard

Btw, did anyone notice they removed access to "who starred the project" recently. You can only check your own repos, not external ones. It was very nice to find related people before.

From the sources it is based on BL606P

https://en.bouffalolab.com/product/?type=detail&id=16

voice processing is in hardware unfortunately, but it exposes some things like DOA

You'd better use offline stt/tts for this.

Kroko models are more accurate and their size is just a hundred megabytes compared to parakeet (2.5 gigabytes in default fp32)

> Running a family was a brutal two-person job -- and the kids had to dive in to help out the second they could lift something heavier than a couple pounds.

Orphanes did struggle but most families were not just two person, families were big and supported by community.

You can check whale sound recognition project https://arxiv.org/abs/2104.08614

And moreover, you can not tune those models for practical applications. The model is originally trained on very clean data, so lower layers are also not very stable for diverse inputs. To finetune you have to update the whole model, not just upper layers.

This model is actually expected to be bad for popular languages, just like previous MMS it is not accurate at all, it wins by supporting something rare well but never had good ASR accuracy even for Swedish etc. It is more a research thing than a real tool. Unlike Whisper.

It is useless actually. Very slow and quality is suboptimal and it is just speech generation component. See discussion here:

https://github.com/SesameAILabs/csm/issues/80

No, there are mathematical reasons LLMs are better. They are trained with multiobjective loss (coding skills, translation skills, etc) so they understand the world much better than MLM. Original post discuss that but with more words and points than necessary.

It is actually pretty straightforward why those model "reason" or, to be more exact, can operate on a complex concepts. By processing huge amount of texts they build an internal representation where those concepts are represented as a simple nodes (neurons or groups). So they really distill knowledge. Alternatively you can think about it as a very good principal component analysis that can extract many important aspects. Or like a semantic graph built automatically.

Once knowledge is distilled you can build on top of it easily by merging concepts for example.

So no secret here.

There is also a glitch in "dialogue"

Anyone except me thinks he doesn't look very healthy? Its strange he is kind of slow on the video where he enters the room. Maybe some biohacking.

Yes, it is one of the important aspects. In particular if you use TTS to create an audiobook or in a video production.

Err, I deeply respect Amazon TTS team but this paper and synthesis is..... You publish the paper in 2024 and include YourTTS in your baselines to look better. Come on! There is XTTS2 around!

Voice sounds robotic and plain. Most likely a lot of audiobooks in training data and less conversational speech. And dropping diffusion was not a great idea, voice is not crystal clear anymore, it is more like a telephony recording.

Metavoice is one of a dozen GPT-based TTS systems around starting from Tortoise. And not that great honestly. You can clearly hear "glass scratches" in their sound, it is because they trained on MP3-compressed data.

There are much more clear sounding systems around. You can listen for StyleTTS2 to compare.

Good improvements for many languages, numbers here

https://github.com/openai/whisper/blob/main/language-breakdo...

Ok, first we screwed buffers by making them globally tracked instead of just a piece of memory. Now its time to break all binary modules again.

Ok, but the photos look very suspicious. 1400 year gold right from the ground shouldn't shine like that. Compare to the coins here for example

https://www.smithsonianmag.com/smart-news/ancient-welsh-gold...

Great thanks a lot.

So we have numbers on PTB original perplexity 8.79 quantized 9.68, already 10% worse. And PPL reported per token I suppose? Because word PPL for PTB must be around 20, not less than 10.

Any numbers on more complex tasks then? like QA?

Do you have the numbers? I suspect is is way worse. Original llama.cpp authors never measure any numbers as well.

It is not really llama, it is llama quantized to 4bit. Not even the quality of original 7B. I could also quantize it to 1 bit and claim it runs on my RPI3.

In such an actively developed area like TTS/ASR there is high chance that custom solution would fit your needs much better. The feature set of TTS is actually pretty large and hard to combine in a single ML model. No free lunch you know.

For example if you look for singing voice, they might suggest you an adapted model that is good specifically for singing.

The testing process is also not very straight, you need to understand what to test and how to test properly. For example, some of their voices might be better for questions, some for news.

You'd better talk to them.

Vosk

https://alphacephei.com/vosk/lm

You can restrict the vocabulary the way you like, for example, here is the chess app built with Vosk

https://www.chessvis.com/

Thats probably medium model

Do I understand correctly banana pricing is that it costs $1.87 per hour, so the hour of audio with large model costs you about $1? Thats probably a bit too expensive compared to cloud providers.

https://podscript.ai/ too!

There is new better one https://github.com/rhasspy/larynx2

Picovoice technology is really nice! I happily recommend it to clients looking for lightweight ASR.

You might not understand but there is a huge amount of work behind this simple demo.