Questions about model inference

[gdpgxy]

Hello! This is very impressive work!

I have a few questions regarding the model's bitrate:

1. If I want to achieve a bitrate of 0.125kbps, would it be sufficient to configure the decoder to only use the first two layers of the RVQ, like this:

dec_rvq2 = model.decode(enc.audio_codes[:2], return_dict=True)
wav_rvq2 = dec_rvq2.audio.squeeze(0)
torchaudio.save("demo/demo_rec_rvq2.wav", wav_rvq2, sample_rate=model.sampling_rate)

2. For streaming inference, what is the minimum processing latency required for the model? Specifically, for frame-by-frame inference, how many past frames need to be considered? This is similar to feeding the current frame as input while the model is still processing previous frames—the inference for the current frame cannot be completed until the results of the preceding frames are available. This can be understood as a fixed algorithmic latency inherent to the model architecture, independent of the inference platform.

Looking forward to your response!
Best regards!

[alpacaking]

Thank you for your interest in MOSS-Audio-Tokenizer.

1. MOSS-Audio-Tokenizer uses 32 RVQ layers in total, with a full bitrate of 4 kbps. Therefore, 0.125 kbps corresponds to using only the first RVQ layer. Using the first two RVQ layers would correspond to 0.25 kbps instead of 0.125 kbps.

2. For streaming inference, the theoretical minimum algorithmic latency of our model is one frame. Since the model operates at 12.5 frames per second, the minimum latency is 0.08 seconds (80 ms).
   For frame-by-frame inference, our model is strictly causal and does not rely on any future frames. In other words, the current frame can be processed once the current 80 ms frame is available. If your question is about historical context, each causal transformer block uses a context duration of up to 10 seconds.

If you are interested, you may refer to our paper for more details. https://arxiv.org/abs/2602.10934

[gdpgxy]

Thank you for your interest in MOSS-Audio-Tokenizer.

1. MOSS-Audio-Tokenizer uses 32 RVQ layers in total, with a full bitrate of 4 kbps. Therefore, 0.125 kbps corresponds to using only the first RVQ layer. Using the first two RVQ layers would correspond to 0.25 kbps instead of 0.125 kbps.
2. For streaming inference, the theoretical minimum algorithmic latency of our model is one frame. Since the model operates at 12.5 frames per second, the minimum latency is 0.08 seconds (80 ms).
   For frame-by-frame inference, our model is strictly causal and does not rely on any future frames. In other words, the current frame can be processed once the current 80 ms frame is available. If your question is about historical context, each causal transformer block uses a context duration of up to 10 seconds.

If you are interested, you may refer to our paper for more details. https://arxiv.org/abs/2602.10934

Thank you very much for your patient reply!
I will read your paper in detail.
Best regards!