docs/architecture.md

Mars 5 technical details

While we do not have the time for a proper full writeup of the details of Mars5, its design, training, and implementation, we at least try give a more detailed overview here of how Mars5 works.

hubconf object/api

After loading the model with torch.hub.load, two objects are returned, a Mars5TTS, and the dataclass of the inference config to use when calling the mars5.tts() method. Concretely, the main methods of the mars5 object are:

# The init function, called automatically when you initialize the 
# model from torch.hub.load(). If you want, you can pass in your
# own custom checkpoints here to initalize the model with your 
# own model, tokenizer, etc...
def __init__(self, ar_ckpt, nar_ckpt, device: str = None) -> None:
    # ... initialization code ...

# Main text-to-speech function, converting text and a reference
# audio to speech. 
def tts(self, text: str, ref_audio: Tensor, ref_transcript: str | None, 
            cfg: InferenceConfig) -> Tensor:
        """ Perform TTS for `text`, given a reference audio `ref_audio` (of shape [sequence_length,], sampled at 24kHz) 
        which has an associated `ref_transcript`. Perform inference using the inference 
        config given by `cfg`, which controls the temperature, top_p, etc...
        Returns:
        - `ar_codes`: (seq_len,) long tensor of discrete coarse code outputs from the AR model.
        - `out_wav`: (T,) float output audio tensor sampled at 24kHz.
        """

# Utility function to vocode encodec tokens, if one wishes 
# to hear the raw AR model ouput by vocoding the `ar_codes` 
# returned above.
def vocode(self, tokens: Tensor) -> Tensor:
    """ Vocodes tokens of shape (seq_len, n_q) """

Model design

Mars 5 follows a two-stage AR-NAR design according to the diagram on the main page.

AR component

The AR model follows a Mistral-style decoder-only transformer model to predict Encodec L0 codes (the lowest/most coarse level quantization codes). Overall, the AR and NAR model is going to predict all 8 codebook entries of the Encodec 6kbps codec. The AR model design is given below:

Figure: autoregressive component of Mars 5. During training, the initial 6kbps encodec tokens of the speech are fed through a small encoder-only transformer, producing a single output vector corresponding to an implicit speaker embedding. This vector is concatenated with learnt embeddings corresponding to the text tokens, and L0 speech tokens, after byte-pair encoding tokenization.

The AR model is trained using the standard next-token prediction task of language models with a cross-entropy loss with the next token, given a smaller weight to text tokens. During inference, we iteratively sample from the transformer to produce the desiged L0 codes. When we use a shallow clone, then the reference audio is fed into the transcript to make the implicit speaker embedding used in the input sequence. When we use a deep clone, the above is done, but we also concatenate the reference transcript with the desired text, and the reference audio tokens with the input sequence before we start sampling the output. In pseudocode:

speaker_embedding <- speaker_conditioning_transformer(ref audio)
if deep_clone:
    prompt = concatenate( speaker embedding, reference text, target text, reference L0 speech codes )
else:
    prompt = concatenate( speaker embedding, target text )

ar output <- autoregressively sample from prompt

While a deep clone provides a more accurate cloning of the reference speaker identity and prosody, it requires knowledge of the reference transcript and takes longer to do inference.

NAR component

After the AR model has predicted the L0 encodec codes, we need a way to predict the remaining 7 codebooks of the 6kbps Encodec codec. This is what the NAR model is trained to do, using a multinomial diffusion framework. Concretely, the diffusion process is a discrete DDPM, whereby at each timestep in the diffusion process, it takes in a sequence of (batch size, sequence length, n_codebooks) and produces an output categorical distribution over each codebook, i.e. an output of shape (batch size, sequence length, n_codebooks, 1024), since each encodec codebook has 1024 possible values. The architecture of the model looks as follows:

Figure: Mars 5 non-autoregressive component. It follows an encoder-decoder transformer architecture, whereby the encoder computes an implicit speaker embedding like the AR model, and concatenates that along with the target to form an input sequence to a transformer encoder. The transformer decoder predicts the distribution of all 8 encodec codebook tokens given a partly noised input at some diffusion timestep t.

The encoder and decoder transformers are simple nn.Transformer variants with sinusoidal positional embeddings and SwiGLU activations. A multinomial diffusion manager controls the forward and reference diffusion processes during inference and training according to a cosine diffusion schedule. Diffusion is performed independently of the sequence length or codebook index.

During training and inference, the L0 codebooks of the input at timestep $t$ are overridden (i.e. not noised in the forward diffusion process) with either the ground truth L0 codes (during training) or the AR model's predictions (during inference). Like the AR model, the NAR model can perform inference in either a shallow clone way or a deep clone way. And, like the AR model, the difference between the two is, with a deep clone, we concatenate the reference text to the input text sequence, and the reference speech codes (the full values for all 8 codebooks) to the decoder input sequence $x$. During inference, we then treat the portion of $x$ corresponding to the reference codec codes, and all the AR L0 codes, as 'fixed' and effectively perform diffusion inpainting for the remaining missing codec codes. The figure below explains what the input to the decoder looks like for a deep clone:

This allows us to use diffusion inpainting techniques like RePaint to improve the quality of the output at the cost of more inference time. We've implemented this in the the diffusion config used in the NAR inference code (see it here), and you can simply increase the jump_len and jump_n_sample to greater than 1 to use RePaint inpainting to improve NAR performance.