add

.gitattributes

@@ -41,6 +41,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.npz filter=lfs diff=lfs merge=lfs -text
 *.pickle filter=lfs diff=lfs merge=lfs -text
 *.pkl filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
 *.tar filter=lfs diff=lfs merge=lfs -text
 *.wasm filter=lfs diff=lfs merge=lfs -text
 *.wav filter=lfs diff=lfs merge=lfs -text

README.md

@@ -21,7 +21,7 @@ This is the code for the 48khz stereo version of MOSS-Audio-Tokenizer presented
 **Key Features:**
 
 *   **Extreme Compression & Variable Bitrate**: It compresses 48kHz stereo audio into a remarkably low frame rate of 12.5Hz. Utilizing a 32-layer Residual Vector Quantization stack, it supports high-fidelity reconstruction across a wide range of bitrates.
-*   **Pure Transformer Architecture**: The model features a "CNN-free" homogeneous architecture built entirely from Causal Transformer blocks. With 1.6B combined parameters (Encoder + Decoder), it ensures exceptional scalability and supports low-latency streaming inference.
+*   **Pure Transformer Architecture**: The model features a "CNN-free" homogeneous architecture built entirely from Causal Transformer blocks. With 2B combined parameters (Encoder + Decoder), it ensures exceptional scalability and supports low-latency streaming inference.
 *   **Large-Scale General Audio Training**: Trained on 3 million hours of diverse audio data, the model excels at encoding and reconstructing all audio domains, including speech, sound effects, and music.
 *   **Unified Semantic-Acoustic Representation**: While achieving state-of-the-art reconstruction quality, Cat produces discrete tokens that are "semantic-rich," making them ideal for downstream tasks like speech understanding (ASR) and generation (TTS).
 *   **Fully Trained From Scratch**: Cat does not rely on any pretrained encoders (such as HuBERT or Whisper) or distillation from teacher models. All representations are learned autonomously from raw data.
@@ -34,55 +34,12 @@ This repository contains a lightweight remote-code implementation that mirrors t
 `transformers.models.moss_audio_tokenizer` module. It is hosted as a Hugging Face Hub model repository and should be
 loaded with `trust_remote_code=True`.
 
-## Model Details
-
-- **Architecture:** Cat (Causal Audio Tokenizer with Transformer), a CNN-free neural audio codec/tokenizer.
-- **Sampling rate:** 48 kHz.
-- **Channels:** stereo public waveform interface.
-- **Token frame rate:** 12.5 Hz.
-- **Quantization:** 32-layer residual vector quantization stack.
-- **Checkpoint size:** the safetensors index reports 2,123,701,248 total parameters.
-- **Weight format:** sharded `safetensors` weights with a `model.safetensors.index.json` index.
-
-## Intended Use
-
-MOSS-Audio-Tokenizer-v2 is intended for research and development on audio tokenization, neural codec reconstruction,
-native audio foundation models, speech/audio understanding, speech generation, and related downstream modeling. It can
-encode 48 kHz stereo waveforms into discrete audio codes and decode those codes back to waveforms.
-
-This model is not intended for use in applications that impersonate a real person, reproduce private or copyrighted
-audio without permission, or make high-stakes decisions from reconstructed audio without additional validation.
-
-## Training Data And Procedure
-
-The model was trained from scratch on 3 million hours of diverse audio data, covering speech, sound effects, and music,
-as described in the accompanying paper. The training pipeline jointly optimizes the encoder, quantizer, decoder,
-discriminator, and a decoder-only LLM used for semantic alignment.
-
-The full training data mixture is not included in this repository. For details on dataset composition, filtering, and
-training/evaluation methodology, refer to the paper.
-
-## Evaluation
-
-The model is designed to provide high-fidelity reconstruction and semantically rich discrete representations across
-speech, sound effects, and music. Please refer to the paper for the full benchmark setup and quantitative results.
-
-## Limitations
-
-- Audio outside the 48 kHz stereo setting may require resampling and channel conversion before inference.
-- Reconstruction quality depends on audio domain, signal quality, selected number of RVQ layers, and inference settings.
-- The repository uses custom Transformers remote code, so users should review the code and pin a trusted revision in
-  production deployments.
-- `flash_attention_2` is optional; if it is unavailable, use the default `sdpa` attention implementation.
-
-## Requirements
-
-- Python 3.10 or newer.
-- PyTorch.
-- Transformers. This checkpoint was prepared with `transformers_version` set to `4.56.0.dev0`; use a recent Transformers
-  build that supports custom remote-code models.
-- `torchaudio` for the examples below.
-- Optional: `flash-attn` if using `model.set_attention_implementation("flash_attention_2")`.
+<br>
+<p align="center">
+    <img src="images/arch.png" width="95%"> <br>
+    Architecture of MossAudioTokenizer
+</p>
+<br>
 
 ## Usage