Duplicate from HKUSTAudio/xcodec2

Co-authored-by: HKUST Audio <HKUST-Audio@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,69 @@
+---
+license: cc-by-nc-4.0
+tags:
+- audio-to-audio
+pipeline_tag: audio-to-audio
+---
+
+[![arXiv](https://img.shields.io/badge/arXiv-Paper-<COLOR>.svg)](https://arxiv.org/abs/2502.04128)
+**Update (2025-02-13):** Add [Llasa finetune instruction](https://github.com/zhenye234/LLaSA_training/tree/main/finetune).
+
+**Update (2025-02-07):** Our paper has been released!
+
+
+## Paper
+ 
+LLaSA: Scaling Train Time and Inference Time Compute for LLaMA based Speech Synthesis  
+
+Codec Does Matter: Exploring the Semantic Shortcoming of Codec for Audio Language Model (AAAI 2025, xcodec 1.0)
+
+
+# Getting Started with XCodec2 on Hugging Face
+XCodec2 is a speech tokenizer that offers the following key features:
+
+1. **Single Vector Quantization**
+2. **50 Tokens per Second**
+3. **Multilingual Speech Semantic Support and High-Quality Speech Reconstruction**
+
+
+To use `xcodec2`, ensure you have it installed. You can install it using the following command:
+
+```bash
+conda create -n xcodec2 python=3.9
+conda activate xcodec2
+pip install xcodec2  (Use `xcodec2==0.1.5` for codec inference and llasa fine-tuning. I’ve removed unnecessary dependencies, and it works fine in my testing. However,  I’m not sure if other problems may arise. If you prefer more stability, I recommend using `xcodec2==0.1.3` which accurately aligns during my codec training.)
+ 
+```
+Then,
+```python
+import torch
+import soundfile as sf
+from transformers import AutoConfig
+
+ 
+from xcodec2.modeling_xcodec2 import XCodec2Model
+ 
+model_path = "HKUSTAudio/xcodec2"  
+ 
+model = XCodec2Model.from_pretrained(model_path)
+model.eval().cuda()   
+
+ 
+wav, sr = sf.read("test.wav")   
+wav_tensor = torch.from_numpy(wav).float().unsqueeze(0)  # Shape: (1, T)
+
+ 
+with torch.no_grad():
+   # Only 16khz speech
+   # Only supports single input. For batch inference, please refer to the link below.
+    vq_code = model.encode_code(input_waveform=wav_tensor)
+    print("Code:", vq_code )  
+
+    recon_wav = model.decode_code(vq_code).cpu()       # Shape: (1, 1, T')
+
+ 
+sf.write("reconstructed.wav", recon_wav[0, 0, :].numpy(), sr)
+print("Done! Check reconstructed.wav")
+```
+
+# If you want to train your own xcodec2, batch inference, or large-scale code extraction, the code is released [here](https://github.com/zhenye234/X-Codec-2.0).

ckpt/epoch=4-step=1400000.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02d8c58b95ac84de701a94e351b6981523bd106cbd05863a09cc56dd148c689e
+size 8306367217

config.json

@@ -0,0 +1,11 @@
+{
+    "model_type": "xcodec2",
+    "semantic_hidden_size": 1024,
+    "codec_encoder_hidden_size": 1024,
+    "codec_decoder_hidden_size": 1024,
+    "use_vocos": true,
+    "architectures": [
+      "XCodec2Model"
+    ]
+  }
+  

configuration_bigcodec.py

@@ -0,0 +1,19 @@
+from transformers import PretrainedConfig
+
+class BigCodecConfig(PretrainedConfig):
+    model_type = "bigcodec"
+
+    def __init__(
+        self,
+        # 下面这些只是示例超参
+        semantic_hidden_size=1024,
+        codec_encoder_hidden_size=1024,
+        codec_decoder_hidden_size=1024,
+        use_vocos=True,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.semantic_hidden_size = semantic_hidden_size
+        self.codec_encoder_hidden_size = codec_encoder_hidden_size
+        self.codec_decoder_hidden_size = codec_decoder_hidden_size
+        self.use_vocos = use_vocos

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a27b1544a394b9aff722da55366281f00480a5479fe5cbfc84f2e2e74eab4db0
+size 3291104560

modeling_xcodec2.py

@@ -0,0 +1,165 @@
+import torch
+import torch.nn as nn
+from transformers import PreTrainedModel
+from configuration_bigcodec import BigCodecConfig
+
+# 请确保这些模块路径是正确的
+from vq.codec_encoder import CodecEncoder_Transformer
+from vq.codec_decoder_vocos import CodecDecoderVocos
+from vq.module import SemanticEncoder
+from transformers import AutoFeatureExtractor, Wav2Vec2BertModel
+
+class XCodec2Model(PreTrainedModel):
+    config_class = BigCodecConfig
+
+    def __init__(self, config: BigCodecConfig):
+        super().__init__(config)
+
+        # 1) 语义模型
+        self.semantic_model = Wav2Vec2BertModel.from_pretrained(
+            "facebook/w2v-bert-2.0",
+            output_hidden_states=True
+        )
+        self.semantic_model.eval()
+
+        self.SemanticEncoder_module = SemanticEncoder(
+             config.semantic_hidden_size,
+             config.semantic_hidden_size,
+             config.semantic_hidden_size
+        )
+
+        # 2) Codec Encoder
+        self.CodecEnc = CodecEncoder_Transformer()
+
+        # 3) Codec Decoder
+        self.generator = CodecDecoderVocos()  
+
+        # 4) 两个全连接层
+        self.fc_prior = nn.Linear(2048, 2048)
+        self.fc_post_a = nn.Linear(2048, 1024)
+        feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/w2v-bert-2.0")
+        self.feature_extractor = feature_extractor
+
+    def forward(self, input_waveform, sample_rate=16000):
+        """
+        这里的 forward 不一定要叫 forward，也可以拆成别的方法；
+        但是如果想兼容 pipeline，需要在 forward 里给出核心逻辑。
+        
+        参数：
+          input_waveform: [batch_size, waveform_length]
+          sample_rate: 默认 16000
+        返回：
+          重构后的语音音频 (Tensor)
+        """
+        # 1) 特征提取
+        # 如果需要 padding，可以在这里做
+        input_features = self.feature_extractor(
+            input_waveform, 
+            sampling_rate=sample_rate, 
+            return_tensors="pt"
+        ).input_features.to(self.device)  # [batch, frames, feat_dim]
+
+        # 2) 语义层
+        semantic_output = self.semantic_model(input_features)
+        semantic_hidden_16 = semantic_output.hidden_states[16]  # 取第16层
+        semantic_hidden_16 = semantic_hidden_16.transpose(1, 2)  # [batch, hidden_dim, frames]
+        semantic_encoded = self.SemanticEncoder_module(semantic_hidden_16)
+
+        # 3) codec encoder
+        wav = input_waveform.unsqueeze(1).to(self.device)  # shape: [batch, 1, time]
+        vq_emb = self.CodecEnc(wav)  # [batch, time//down, 1024] 只是示例
+        vq_emb = vq_emb.transpose(1, 2)  # -> [batch, 1024, frames]
+
+        # 对齐语义向量的时间帧数，这里只做示例处理
+        # 真实做法里可能要先对齐维度
+        if vq_emb.shape[-1] != semantic_encoded.shape[-1]:
+            # 简单强行截断或补零都行，需要你自己决定
+            min_len = min(vq_emb.shape[-1], semantic_encoded.shape[-1])
+            vq_emb = vq_emb[:, :, :min_len]
+            semantic_encoded = semantic_encoded[:, :, :min_len]
+
+        # 4) 拼接
+        concat_emb = torch.cat([semantic_encoded, vq_emb], dim=1)  # [batch, 1024 + 1024, frames]
+
+        # 5) fc_prior
+        concat_emb = self.fc_prior(concat_emb.transpose(1, 2)).transpose(1, 2)
+
+        # 6) decoder 的量化部分
+        _, vq_code, _ = self.generator(concat_emb, vq=True)
+        vq_post_emb = self.generator.quantizer.get_output_from_indices(vq_code.transpose(1, 2))
+        vq_post_emb = vq_post_emb.transpose(1, 2)
+
+        # 7) fc_post_a
+        vq_post_emb = self.fc_post_a(vq_post_emb.transpose(1, 2)).transpose(1, 2)
+
+        # 8) 最后解码成波形
+        recon_audio = self.generator(vq_post_emb.transpose(1, 2), vq=False)[0]
+        # recon_audio: [batch, time]
+        return recon_audio
+
+    def encode_code(self, input_waveform, sample_rate=16000):
+        """
+        将输入的音频编码为代码表示。
+
+        参数：
+          input_waveform: [batch_size, waveform_length]
+          sample_rate: 默认 16000
+        返回：
+          编码后的代码 (Tensor)
+        """
+        with torch.no_grad():
+            # 1) 特征提取
+            input_features = self.feature_extractor(
+                input_waveform, 
+                sampling_rate=sample_rate, 
+                return_tensors="pt"
+            ).input_features.to(self.device)  # [batch, frames, feat_dim]
+
+            # 2) 语义层
+            semantic_output = self.semantic_model(input_features)
+            semantic_hidden_16 = semantic_output.hidden_states[16]  # 取第16层
+            semantic_hidden_16 = semantic_hidden_16.transpose(1, 2)  # [batch, hidden_dim, frames]
+            semantic_encoded = self.SemanticEncoder_module(semantic_hidden_16)
+
+            # 3) codec encoder
+            wav = input_waveform.unsqueeze(1).to(self.device)  # shape: [batch, 1, time]
+            vq_emb = self.CodecEnc(wav)  # [batch, time//down, 1024] 只是示例
+            vq_emb = vq_emb.transpose(1, 2)  # -> [batch, 1024, frames]
+
+            # 对齐语义向量的时间帧数，这里只做示例处理
+            if vq_emb.shape[-1] != semantic_encoded.shape[-1]:
+                min_len = min(vq_emb.shape[-1], semantic_encoded.shape[-1])
+                vq_emb = vq_emb[:, :, :min_len]
+                semantic_encoded = semantic_encoded[:, :, :min_len]
+
+            # 4) 拼接
+            concat_emb = torch.cat([semantic_encoded, vq_emb], dim=1)  # [batch, 2048, frames]
+
+            # 5) fc_prior
+            concat_emb = self.fc_prior(concat_emb.transpose(1, 2)).transpose(1, 2)
+
+            # 6) decoder 的量化部分，获取code
+            _, vq_code, _ = self.generator(concat_emb, vq=True)
+            # vq_code: [batch, frames]
+            return vq_code
+
+    def decode_code(self, vq_code):
+        """
+        将编码后的代码解码回音频。
+
+        参数：
+          vq_code: 编码后的代码 (Tensor) [batch, frames]
+        返回：
+          解码后的音频 (Tensor) [batch, waveform_length]
+        """
+        with torch.no_grad():
+            # 获取量化后的嵌入
+            vq_post_emb = self.generator.quantizer.get_output_from_indices(vq_code.transpose(1, 2))
+            vq_post_emb = vq_post_emb.transpose(1, 2)  # [batch, 1024, frames]
+
+            # 7) fc_post_a
+            vq_post_emb = self.fc_post_a(vq_post_emb.transpose(1, 2)).transpose(1, 2)  # [batch, 1024, frames]
+
+            # 8) 最后解码成波形
+            recon_audio = self.generator(vq_post_emb.transpose(1, 2), vq=False)[0]  # [batch, time]
+            return recon_audio

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8cb939062f3930e56ff22082f49c95461aedc8ceade7ff7b16a1b10f1e92e0be
+size 3291343655

test.py

@@ -0,0 +1,21 @@
+import torch
+import soundfile as sf
+from transformers import AutoConfig
+
+from  modeling_xcodec2 import XCodec2Model
+
+model_path = "/data/zheny/xcodec2"  # 这是你在 huggingface 上的仓库名
+
+model = XCodec2Model.from_pretrained(model_path)
+model.eval().cuda()
+
+# 准备一段音频
+wav, sr = sf.read("test.flac")
+wav_tensor = torch.from_numpy(wav).float().unsqueeze(0)  # [1, time]
+
+with torch.no_grad():
+    vq_code = model.encode_code(input_waveform=wav_tensor )
+    print(vq_code)
+    recon_wav = model.decode_code(vq_code).cpu()
+
+sf.write("reconstructed.wav", recon_wav[0,0,:].numpy(), sr)

vq/__init__.py

@@ -0,0 +1,4 @@
+from vq.codec_encoder import CodecEncoder
+from vq.codec_decoder import CodecDecoder
+from vq.codec_decoder_vocos import CodecDecoderVocos
+from vq.codec_encoder import CodecEncoder_Transformer,CodecEncoder_only_Transformer