Llasa: Llama-Based Speech Synthesis | Hacker News Reader

Llasa: Llama-Based Speech Synthesis

llasatts.github.io

168 pointsCalmStorm22 comments

Open articleView on HN

Odd that the page doesn't seem to link to either,

paper: https://arxiv.org/abs/2502.04128

github: https://github.com/zhenye234/LLaSA_training

Interesting that there isn't a mention of Orpheus as prior art either since it's the exact same thing.

(https://github.com/canopyai/Orpheus-TTS)

> Interesting that there isn't a mention of Orpheus as prior art either

Llasa-3b (https://huggingface.co/HKUSTAudio/Llasa-3B) came out before Orpheus (https://huggingface.co/canopylabs/orpheus-3b-0.1-ft).

> it's the exact same thing.

They're very similar, but they're not the exact same thing.

Llasa uses xcodec2, a much simpler, lossless 16khz wav codec. This makes it superior for one-shot voice cloning.

Orpheus' 24khz snac codec is lossy which makes it difficult to use for zero-shot cloning as the reference audio gets degraded during tokenization. You can test this here: https://huggingface.co/spaces/Gapeleon/snac_test

But when finetuned on 50+ audio samples, it produces much cleaner 24khz audio than Llasa, and the snac model is much easier to run on consumer hardware than xcodec2 (87t/s for realtime speech, which can be achieved on an RTX3080 for example)

Do you happen to know why Orpheus and Llasa use Finetuning for voice cloning?

Zonos uses 128-float embeddings for voices and it seems so much nicer. Because you can just mix and match voices without changing the model.

No, you just condition it with text-voice token pairs and then when conditioning further inference w/ text the voice tokens tend to match the pairs further up in the context.

Isn't xcodec2 also lossy? I thought it is also just another neural codec (50 tok/s, single codebook).

What are people using to upsampling back to 44,1 or 48 khz? Anything fancy?

They’re both lossy. They use a VAE-VQ type architecture trained with a combination of losses/discriminators. The differences are mainly the encoder/decoder architecture, the type of bottleneck quantization (RVQ, FSQ, etc.) and of course the training data.

LLaSA is a simple framework for speech synthesis that employs a single-layer vector quantizer (VQ) codec and a single Transformer architecture to fully align with standard LLMs such as LLaMA.

Probably the title should have the correct capitalization then. Cause I was fully expecting a speech synthesis tool that sounded like llamas talking human language and now I'm bummed out!

> employs a single-layer vector quantizer (VQ) codec and a single Transformer architecture to fully align

I really wish when new models were released that they would draw a diagram of all the layers and the tensor input and output sizes at each layer, with zoom in/out capabilities if needed using D3.js or whatever visualization framework if needed. Every single layer should be on there with its input and output sizes.

These one-sentence descriptions, and approximate block diagrams with arrows pointing at each other are never enough to understand how something is actually implemented.

This already exists in Transformer Lab and ONNX (not recommended for transformers).

You can also build a custom version of llama.cpp that writes out the ggml compute graph. What's irritating is that hugging face didn't add it to their GGUF file viewer.

Oh, sure, for the well-known models that are already on there.

I just wish that new research would always spell it out in full instead of these silly block diagrams labelled with just e.g. "Cross Attention" and not the exact parameters, number of heads, layer sizes, etc.

Also some of these diagrams use a + for concatenation and some use it for addition, that's another headache to figure out, having layer sizes would make it clear.

That might be intentional.

Sounds like a solid SaaS business plan!

I can't wait see this integrated into Open WebUI! These sound amazing.

You can run an openai-compatible endpoint and point open-webui at it if you want this. I had to add a function to filter out markdown lists, code, etc as the model was choking on them.

the long 'uuuuhhhhhhh' from some of the lesser models is killing me.

This finetune seems pretty stable (1b llasa) https://huggingface.co/spaces/HKUST-Audio/Llasa-1B-multi-spe...

1B is actually huge for a TTS model. Here's an 82m model with probably the most stable/coherent output of all the open weights tts models I've tested: https://huggingface.co/spaces/hexgrad/Kokoro-TTS

But if you mean zero-shot cloning, yeah they all seem to have those slurred speech artefacts from time to time.

based on the samples, it really seams like anything smaller than 3B is pretty useless.

If you're doing a home lab voice assistant 1B is nice, because on a 12gb gpu you can run a moderately competent 7b LLM and two 1b models; 1 for speech to text and also text to speech, plus some for the wake word monitor. Maybe in a couple of years we can combine all this into a single ~8b model that runs efficiently on 12gb gpu. Nvidia doesn't seem very incentivized right now to sell consumer GPUs that can run all this on a single consumer grade chip when they're making so much money selling commercial grade 48gb cards.

Hui for the activation word?

Shouldn't there be some hardware module be available similar to how Alexa, Siri and Google do it?

Whith a ring buffer detection the word without recording everything?

the mispronunciation of 行 and 行 in the Chinese sample is killing me too XD

Reply on news.ycombinator.com