This generic answer from Wikipedia is not very helpful in this context. Zero-shot voice cloning in TTS usually means that data of the target speaker you want the generated speech to sound like does not need to be included in the training data used to train the TTS models. In other words, you can provide an audio sample of the target speaker together with the text to be spoken to generate the audio that sounds like it was spoken by that speaker.
Zero-shot means zero-retraining, so think along the lines of "Do you need to modify the weights? Or can you keep the weights fixed and you only need to supply an example?"
> Zero-shot means zero-retraining, so think along the lines of "Do you need to modify the weights? Or can you keep the weights fixed and you only need to supply an example?"
I would caution that using the term "example" suggests further learning happens at inference-time, which isn't the case.
For LLMs, the entire prompt is the input and conveys both the style and the content vectors. In zero-shot voice cloning, we provide the exact same inputs vectors but just decoupled. Providing reference audio is no different than including "Answer in the style of Sir Isaac Newton" in an LLM's prompt. The model doesn't 'learn' the voice; it simply applies the style vector to the content during the forward pass.
Providing inference-time context (in this case, audio) is no different than giving a prompt to an LLM. Think of it as analogous to an AGENTS.md included in a prompt. You're not retraining the model, you're simply putting the rest of the prompt into context.
If you actually stopped and fine-tuned the model weights on that single clip, that would be one-shot learning.
In the olden days of 2023, you didn’t just find instruct-tuned models sitting on every shelf.
You could use a base model that has only undergone pretraining and can only generate text continuations based on the input it receives. If you provided the model with several examples of a question followed by an answer, and then provided a new question followed by a blank for the next answer, the model understood from the context that it needed to answer the question. This is the most primitive use of ICL, and a very basic way to achieve limited instruction following behavior.
With this few-shot example, I would call that few-shot ICL. Not zero shot, even though the model weights are locked.
But, I am learning that it is technically called zero shot, and I will accept this, even if I think it is a confusingly named concept.
It’s nonsensical to call it “zero shot” when a sample of the voice is provided. The term “zero shot cloning” implies you have some representation of the voice from another domain - e.g. a text description of the voice. What they’re doing is ABSOLUTELY one shot cloning. I don’t care if lots of STT folks use the term this way, they’re wrong.
I don't disagree, but that's what people started calling it. Zero-shot doesn't make sense anyway, as how would the model know what voice it should sound like (unless it's a celebrity voice or similar included in the training data where it's enough to specify a name).
> Zero-shot doesn't make sense anyway, as how would the model know what voice it should sound like (unless it's a celebrity voice or similar included in the training data where it's enough to specify a name).
It makes perfect sense; you are simply confusing training samples with inference context. "Zero-shot" refers to zero gradient updates (retraining) required to handle a new class. It does not mean "zero input information."
> how would the model know what voice it should sound like
It uses the reference audio just like a text based model uses a prompt.
> unless it's a celebrity voice or similar included in the training data where it's enough to specify a name
If the voice is in the training data, that is literally the opposite of zero-shot. The entire point of zero-shot is that the model has never encountered the speaker before.
With LLMs I've seen zero-shot used to describe scenarios where there's no example, it "take this and output JSON", while one-shot has the prompt include an example like "take this and output JSON, for this data the JSON should look like this".
Thus if you feed a the model target voice, ie an example of the desired output vouce, it sure seems like it should be classified as one-shot.
However it seems the zero-shot in voice cloning is relative to learning, and in contrast to one-shot learning[1].
So a bit overloaded term causing confusion from what I can gather.
The confusion clears up if you stop conflating contextual conditioning (prompting) with actual Learning (weight updates). For LLMs, "few-shot prompting" is technically a misnomer that stuck; you are just establishing a pattern in the context window, not training the model.
In voice cloning, the reference audio is simply the input, not a training example. You wouldn't say an image classifier is doing "one-shot learning" just because you fed it one image to classify. That image is the input. Similarly, the reference audio is the input that conditions the generation. It is zero-shot because the model's weights were never optimized for that specific speaker's manifold.
> So if you get your target to record (say) 1 hour of audio, that's a one-shot.
No, that would still be zero shot. Providing inference-time context (in this case, audio) is no different than giving a prompt to an LLM. Think of it as analogous to an AGENTS.md included in a prompt. You're not retraining the model, you're simply putting the rest of the prompt into context.
If you actually stopped and fine-tuned the model weights on that single clip, that would be one-shot learning.
> Providing inference-time context (in this case, audio) is no different than giving a prompt to an LLM.
Right... And you have 0-shot prompts ("give me a list of animals"), 1-shot prompts ("give me a list of animals, for example: a cat"), 2-shot prompts ("give me a list of animals, for example: a cat; a dog"), etc.
The "shot" refers to how many examples are provided to the LLM in the prompt, and have nothing to do with training or tuning, in every context I've ever seen.
> Right... And you have 0-shot prompts ("give me a list of animals"), 1-shot prompts ("give me a list of animals, for example: a cat"), 2-shot prompts ("give me a list of animals, for example: a cat; a dog"), etc.
> The "shot" refers to how many examples are provided to the LLM in the prompt, and have nothing to do with training or tuning, in every context I've ever seen.
In formal ML, "shot" refers to the number of samples available for a specific class during the training phase. You're describing a colloquial usage of the term found only in prompt engineering.
You can't apply an LLMism to a voice cloning model where standard ML definitions apply.
> This generic answer from Wikipedia is not very helpful in this context.
Actually, the general definition fits this context perfectly. In machine learning terms, a specific 'speaker' is simply a 'class.' Therefore, a model generating audio for a speaker it never saw during training is the exact definition of the Zero-Shot Learning problem setup: "a learner observes samples from classes which were not observed during training," as I quoted.
Your explanation just rephrases the very definition you dismissed.
> That's not what happens in zero-shot voice cloning
It is exactly what happens. You are confusing the task (classification vs. generation) with the learning paradigm (zero-shot).
In the voice cloning context, the class is the speaker's voice (not observed during training), samples of which are generated by the machine learning model.
The definition applies 1:1. During inference, it is predicting the conditional probability distribution of audio samples that belong to that unseen class. It is "predict[ing] the class that they belong to," which very same class was "not observed during training."
Jeez, OP asked what it means in this context (zero-shot voice cloning), where you quoted a generic definition copied from Wikipedia. I defined it concretely for this context. Don't take it as a slight, there is no need to get all argumentative.
Given that the TV is trying to match one digital frame against another digital frame, you could probably get decent results even with something super naive like downsampling to a very low resolution, quantizing the color palette, then looking for a pixel for pixel match.
All this could be done long before any sort of TV-specific image processing, so the only source of "noise" I can think of would be from the various encodings offered by the streaming service (e.g. different resolutions and bitrates). With the right choice of downsample resolution and color quantization I have to imagine you could get acceptable results.
1. disclose my method was not quantifiably measurable as the not model, because that is not important to me, speed of action/development outcomes is more important to me, and because
2. I’ve observed a large gap between benchmark toppers and my own results
But make no mistake, I like have the terminals scrolling live across multiple monitors so I can glance at them periodically and watch their response quality, so I care and notice which give better/worse results.
My biggest goal right now after accuracy is achieving more natural human-like English for technical writing.
Just use something like DSPy/Ax and optimize your module for any given LLM (based on sample data and metrics) and you’re mostly good. No need to manually wordsmith prompts.
reply