Research zipvoice

.gitattributes

@@ -35,16 +35,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 assets/moss_prompts/en_4_4p5s.wav filter=lfs diff=lfs merge=lfs -text
 assets/moss_prompts/zh_1_4p5s.wav filter=lfs diff=lfs merge=lfs -text
-cpp/install/ax650/zipvoice filter=lfs diff=lfs merge=lfs -text
-cpp/install/ax650/zipvoice_axera filter=lfs diff=lfs merge=lfs -text
-cpp/install/axcl/zipvoice filter=lfs diff=lfs merge=lfs -text
-cpp/vocoder/vocos_full.axmodel filter=lfs diff=lfs merge=lfs -text
-cpp/vocoder/vocos_full_ax630c.axmodel filter=lfs diff=lfs merge=lfs -text
-models/axmodels/decoder_part0.axmodel filter=lfs diff=lfs merge=lfs -text
-models/axmodels/decoder_part1.axmodel filter=lfs diff=lfs merge=lfs -text
-models/axmodels/decoder_part2.axmodel filter=lfs diff=lfs merge=lfs -text
-models/axmodels/decoder_part3.axmodel filter=lfs diff=lfs merge=lfs -text
-models/axmodels/encoder.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_ax650/decoder_part0.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_ax650/decoder_part1.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_ax650/decoder_part2.axmodel filter=lfs diff=lfs merge=lfs -text
@@ -54,26 +44,25 @@ models/zipvoice_distill_ax630C/decoder_part0.axmodel filter=lfs diff=lfs merge=l
 models/zipvoice_distill_ax630C/decoder_part1.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax630C/decoder_part2.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax630C/decoder_part3.axmodel filter=lfs diff=lfs merge=lfs -text
-models/zipvoice_distill_ax630C/decoder_part3_0.axmodel filter=lfs diff=lfs merge=lfs -text
-models/zipvoice_distill_ax630C/decoder_part3_1.axmodel filter=lfs diff=lfs merge=lfs -text
-models/zipvoice_distill_ax630C/decoder_part3_2.axmodel filter=lfs diff=lfs merge=lfs -text
-models/zipvoice_distill_ax630C/decoder_part3_3.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax630C/encoder.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax650/decoder_part0.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax650/decoder_part1.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax650/decoder_part2.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax650/decoder_part3.axmodel filter=lfs diff=lfs merge=lfs -text
 models/zipvoice_distill_ax650/encoder.axmodel filter=lfs diff=lfs merge=lfs -text
-outputs/en_long_paragraph_ax650.wav filter=lfs diff=lfs merge=lfs -text
+outputs/en_long_paragraph.wav filter=lfs diff=lfs merge=lfs -text
 outputs/en_long_paragraph_distill_ax650.wav filter=lfs diff=lfs merge=lfs -text
-outputs/en_sentence_ax650.wav filter=lfs diff=lfs merge=lfs -text
+outputs/en_sentence.wav filter=lfs diff=lfs merge=lfs -text
 outputs/en_sentence_distill_ax650.wav filter=lfs diff=lfs merge=lfs -text
-outputs/zh_long_paragraph_ax650.wav filter=lfs diff=lfs merge=lfs -text
+outputs/zh_long_paragraph.wav filter=lfs diff=lfs merge=lfs -text
 outputs/zh_long_paragraph_distill_ax650.wav filter=lfs diff=lfs merge=lfs -text
-outputs/zh_sentence_ax650.wav filter=lfs diff=lfs merge=lfs -text
+outputs/zh_sentence.wav filter=lfs diff=lfs merge=lfs -text
 outputs/zh_sentence_distill_ax650.wav filter=lfs diff=lfs merge=lfs -text
-bin/zipvoice_ax630c filter=lfs diff=lfs merge=lfs -text
-bin/zipvoice_ax650 filter=lfs diff=lfs merge=lfs -text
-bin/zipvoice_axcl filter=lfs diff=lfs merge=lfs -text
-models/vocoder/vocos_full.axmodel filter=lfs diff=lfs merge=lfs -text
-models/vocoder/vocos_full_ax630c.axmodel filter=lfs diff=lfs merge=lfs -text
+en_long_paragraph_ax650.wav filter=lfs diff=lfs merge=lfs -text
+en_long_paragraph_distill_ax650.wav filter=lfs diff=lfs merge=lfs -text
+en_sentence_ax650.wav filter=lfs diff=lfs merge=lfs -text
+en_sentence_distill_ax650.wav filter=lfs diff=lfs merge=lfs -text
+zh_long_paragraph_ax650.wav filter=lfs diff=lfs merge=lfs -text
+zh_long_paragraph_distill_ax650.wav filter=lfs diff=lfs merge=lfs -text
+zh_sentence_ax650.wav filter=lfs diff=lfs merge=lfs -text
+zh_sentence_distill_ax650.wav filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -1,24 +0,0 @@
-# HuggingFace CLI 缓存
-.msc/
-.mv/
-._____temp/
-.cache/
-
-# 运行产物
-outputs/
-reference/
-__pycache__/
-*.pyc
-
-# 编译产物/源码（HUG 仓只放可执行文件，不放源码）
-cpp/src/
-cpp/third_party/
-cpp/toolchains/
-cpp/cmake/
-cpp/scripts/
-cpp/CMakeLists.txt
-cpp/build_*.sh
-cpp/download_bsp.sh
-cpp/*.cpp
-cpp/*.hpp
-build/

README.md

@@ -1,39 +1,18 @@
 # ZipVoice.AXERA
 
-[ZipVoice]((https://github.com/k2-fsa/ZipVoice)) AXERA 板端 & AXCL 算力卡推理 demo。
+[ZipVoice]((https://github.com/k2-fsa/ZipVoice)) AXERA 板端推理 demo。
 
 ## 功能
 
 - 支持中文和英文语音生成。
 - 支持语音克隆。
 - 支持 ZipVoice、ZipVoice Distill
-- 支持 C++ 和 Python 推理
-- 支持 AX650、AX630C
 
 ## 模型说明
 
-ZipVoice Distill 是 ZipVoice 的蒸馏版本，主要优势是在较小性能损失下提升推理速度。
-
-## 性能对比
-
-### Distill 模型 (num_step=4)
-
-| 场景 | 音频时长 | Python AX650<br>耗时 RTF | C++ AX650<br>耗时 RTF | C++ AXCL<br>耗时 RTF | Python AX630C<br>耗时 RTF | C++ AX630C<br>耗时 RTF |
-|------|---------|-------------------------|----------------------|---------------------|--------------------------|-----------------------|
-| 中文句子 | 6.41s | 1.992s/0.311 | 1.057s/0.165 | 0.918s/**0.143** | 10.296s/1.606 | 4.650s/0.725 |
-| 中文段落 | 44.97s | 13.457s/0.301 | 7.357s/0.164 | 6.443s/**0.143** | 71.574s/1.600 | 32.590s/0.725 |
-| 英文句子 | 6.41s | 2.045s/0.319 | 1.067s/0.166 | 0.919s/**0.143** | 10.686s/1.667 | 4.654s/0.726 |
-| 英文段落 | 59.16s | 19.715s/0.305 | 10.529s/0.178 | 9.207s/**0.156** | 106.183s/1.640 | 46.541s/0.787 |
-
-### 普通模型 (num_step=10)
-
-| 场景 | 音频时长 | Python AX650 (耗时/RTF) | C++ AX650 (耗时/RTF) | C++ AXCL (耗时/RTF) |
-|------|---------|------------------------|---------------------|---------------------|
-| 中文句子 | 6.41s | 5.781s / 0.902 | 5.709s / 0.891 | 5.452s / **0.850** |
-| 中文段落 | 44.97s | 40.292s / 0.901 | 39.714s / 0.883 | 38.138s / **0.848** |
-| 英文句子 | 6.41s | 5.711s / 0.891 | 5.693s / 0.888 | 5.447s / **0.850** |
-| 英文段落 | 59.16s | 62.161s / 0.960 | 56.759s / 0.957 | 54.451s / **0.920** |
+ZipVoice Distill 是 ZipVoice 的蒸馏版本，主要优势是在较小性能损失下提升推理速度。初步测试，AX650 ZipVoice Distill 在长文本场景下相比基础版模型约有 3 倍速度提升，RTF 在 0.3 左右，效果没有明显下降。
 
+AX630C 版本当前推理结果差，RTF 约为 1.5 左右，需要继续调优。
 
 ## 模型转换
 
@@ -49,8 +28,6 @@ ZipVoice Distill 是 ZipVoice 的蒸馏版本，主要优势是在较小性能
   - [M4N-Dock（爱芯派Pro）](https://wiki.sipeed.com/hardware/zh/maixIV/m4ndock/m4ndock.html)
   - [M.2 Accelerator Card](https://docs.m5stack.com/zh_CN/ai_hardware/LLM-8850_Card)
 
-- AX630C
-
 
 ## 目录结构
 
@@ -62,30 +39,20 @@ ZipVoice.AXERA/
 ├── models/
 │   ├── zipvoice_ax650/
 │   ├── zipvoice_distill_ax650/
-│   ├── zipvoice_distill_ax630C/
-│   └── vocoder/                  # vocos_full.axmodel（AX650/AX630C）
+│   └── zipvoice_distill_ax630C/
 ├── resources/
-│   ├── vocos-mel-24khz/          # Python 推理用 vocoder 权重
-│   └── zipvoice_hf/              # tokens.txt
-├── bin/                          # C++ 推理可执行文件（源码见 GitHub）
-│   ├── zipvoice_ax650
-│   ├── zipvoice_ax630c
-│   └── zipvoice_axcl
+│   ├── vocos-mel-24khz/
+│   └── zipvoice_hf/
 ├── scripts/
-├── infer_zipvoice_axera.py      # Python 推理入口
-├── run_ax650.sh                 # AX650 板端
-├── run_ax630c.sh                # AX630C 板端
-├── run_axcl.sh                  # AXCL 算力卡
+├── infer_zipvoice_axera.py
 ├── requirements.txt
 └── README.md
 ```
 
-## Python
-
-### 环境
+## 环境
 
 安装 pyaxengine：
-[pyaxengine Releases](https://github.com/AXERA-TECH/pyaxengine/releases/latest) 下载对应版本安装：
+
 ```bash
 pip3 install axengine-x.x.x-py3-none-any.whl
 ```
@@ -98,7 +65,7 @@ conda activate ZipVoice
 pip3 install -r requirements.txt
 ```
 
-### 推理命令
+## 推理命令
 
 进入目录：
 
@@ -106,7 +73,7 @@ pip3 install -r requirements.txt
 cd ZipVoice.AXERA
 ```
 
-#### AX650 ZipVoice
+### AX650 ZipVoice
 
 中文句子：
 
@@ -196,7 +163,7 @@ RTF: 0.9600
 音频：[outputs/en_long_paragraph_ax650.wav](outputs/en_long_paragraph_ax650.wav)  
 提示音：[assets/moss_prompts/en_4_4p5s.wav](assets/moss_prompts/en_4_4p5s.wav)
 
-#### AX650 ZipVoice Distill
+### AX650 ZipVoice Distill
 
 中文句子：
 
@@ -286,44 +253,15 @@ RTF: 0.3045
 音频：[outputs/en_long_paragraph_distill_ax650.wav](outputs/en_long_paragraph_distill_ax650.wav)  
 提示音：[assets/moss_prompts/en_4_4p5s.wav](assets/moss_prompts/en_4_4p5s.wav)
 
-#### AX630C ZipVoice Distill
-
-普通模型推理太慢，这边仅提供 Distill 模型，参数与 AX650 一致，仅 `--model-name` 改为 `zipvoice_distill_ax630C`：
-
-```bash
-python3 infer_zipvoice_axera.py \
-  --model-name zipvoice_distill_ax630C \
-  --text "今天午后天气很好，我打开窗户，听见远处有人聊天，水杯也轻轻晃了一下。" \
-  --prompt-text "不管怎么样我和汤姆还是要感谢贝尔卡金的援手" \
-  --prompt-wav assets/moss_prompts/zh_1_4p5s.wav --seed 42
-```
 
-### 参数说明
+## 参数说明
 
-- `--model-name`：`zipvoice_ax650` / `zipvoice_distill_ax650` / `zipvoice_distill_ax630C`。
+- `--model-name`：选择模型目录。可选 `zipvoice_ax650`、`zipvoice_distill_ax650`、`zipvoice_distill_ax630C`。
 - `--prompt-wav`：参考音频，用于控制音色，建议 3-5s。
 - `--prompt-text`：参考音频对应文本，必须尽量和 `prompt-wav` 内容一致。
 - `--num-step`：采样步数。默认从模型目录的 `runtime_config.json` 读取。
 - `--max-feat-len`：decoder 固定 feature 长度，当前模型均为 1024。
 
-## C++
-
-C++ 推理可执行文件在 `bin/` 下，**无需编译**，直接运行：
-
-```bash
-# AX650 板端
-bash run_ax650.sh distill zh sentence      # distill/standard, zh/en, sentence/paragraph
-# AX630C 板端
-bash run_ax630c.sh zh sentence
-# AXCL 算力卡
-bash run_axcl.sh distill zh paragraph
-```
-
-C++ 源码见 [GitHub: AXERA-TECH/ZipVoice.AXERA/cpp](https://github.com/AXERA-TECH/ZipVoice.AXERA/tree/main/cpp)。
-
-
 ## 参考
 
 - [ZipVoice](https://github.com/k2-fsa/ZipVoice)
-- [Pulsar2 Docs](https://pulsar2-docs.readthedocs.io/en/latest/pulsar2/introduction.html)
-- [模型量化工程](https://github.com/AXERA-TECH/ZipVoice.AXERA)

bin/zipvoice_ax630c

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:adbc16b5766ca2eede9b14b405d5b534e03b66b29043d179b41aa0a5df77d721
-size 527832

bin/zipvoice_ax650

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8653e75af7dc2571df341a6172a2342c19f6c99761ae640261c798c86087fe67
-size 527848

bin/zipvoice_axcl

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:de730337c6c5ecf11e805f7267447ff387cf5ee80131d1ad4875129586285fc2
-size 581640

configuration.json

@@ -1 +0,0 @@
-{"framework": "pytorch", "task": "others", "allow_remote": true}

infer_zipvoice_axera.py

@@ -58,7 +58,6 @@ def parse_args() -> argparse.Namespace:
         help="Allow raw features_len up to this ratio of max_feat_len before splitting",
     )
     p.add_argument("--silence-ms", type=int, default=140)
-    p.add_argument("--vocoder-model", default=None, help="Path to vocos axmodel. If set, use axmodel vocoder instead of PyTorch.")
     return p.parse_args()
 
 
@@ -159,13 +158,7 @@ def main() -> None:
         t_shift=args.t_shift,
     )
     logging.debug("Loading vocoder...")
-    if args.vocoder_model:
-        from scripts.common_infer import load_axmodel_vocoder, axmodel_vocoder_decode
-        vocoder = load_axmodel_vocoder(args.vocoder_model)
-        vocoder_is_axmodel = True
-    else:
-        vocoder = load_vocoder(repo_dir)
-        vocoder_is_axmodel = False
+    vocoder = load_vocoder(repo_dir)
     logging.info("模型加载完成")
 
     audios: list[np.ndarray] = []
@@ -190,16 +183,13 @@ def main() -> None:
             seed=args.seed + index - 1,
         )
 
-        if vocoder_is_axmodel:
-            from scripts.common_infer import axmodel_vocoder_decode
-            audio = axmodel_vocoder_decode(
-                vocoder, pred_features,
-                feat_scale=args.feat_scale, target_rms=args.target_rms, prompt_rms=prompt_rms)
-        else:
-            audio = vocoder_decode_loaded(
-                vocoder, pred_features,
-                feat_scale=args.feat_scale, target_rms=args.target_rms,
-                prompt_rms=prompt_rms)
+        audio = vocoder_decode_loaded(
+            vocoder,
+            pred_features,
+            feat_scale=args.feat_scale,
+            target_rms=args.target_rms,
+            prompt_rms=prompt_rms,
+        )
         segment_sec = time.perf_counter() - t_segment_start
         audio_sec = len(audio) / 24000
 

models/vocoder/vocos_full.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:70a7a897f6beb0b6e8536413a0ea6ee2132bec1257e51208e80eca506bda4c40
-size 15413002

models/vocoder/vocos_full_ax630c.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:4b704aa370e7875099615ebbb2772c24e7d97529df81f11b713ca04e4261903e
-size 16208038

models/zipvoice_distill_ax630C/decoder4_split_manifest.json

@@ -1,104 +1,68 @@
 {
-  "version": 1,
-  "model_type": "zipvoice_distill",
-  "encoder": {
-    "name": "encoder_core",
-    "file": "encoder.axmodel",
-    "inputs": [
-      "cat_tokens"
-    ],
-    "outputs": [
-      "encoded"
-    ]
-  },
-  "decoder_parts": [
-    {
-      "name": "fm_decoder_part0",
-      "file": "decoder_part0.axmodel",
-      "inputs": [
-        "t",
-        "x",
-        "text_condition",
-        "speech_condition",
-        "guidance_scale",
-        "padding_mask"
-      ],
-      "outputs": [
-        "decoder_hidden_p0",
-        "time_hidden"
-      ]
-    },
-    {
-      "name": "fm_decoder_part1",
-      "file": "decoder_part1.axmodel",
-      "inputs": [
-        "decoder_hidden_p0",
-        "time_hidden",
-        "padding_mask"
-      ],
-      "outputs": [
-        "decoder_hidden_p1"
-      ]
-    },
-    {
-      "name": "fm_decoder_part2",
-      "file": "decoder_part2.axmodel",
-      "inputs": [
-        "decoder_hidden_p1",
-        "time_hidden",
-        "padding_mask"
-      ],
-      "outputs": [
-        "decoder_hidden_p2"
-      ]
-    },
-    {
-      "name": "fm_decoder_part3_0",
-      "file": "decoder_part3_0.axmodel",
-      "inputs": [
-        "decoder_hidden_p2",
-        "time_hidden",
-        "padding_mask"
-      ],
-      "outputs": [
-        "/fm_decoder/4/0/bypass/Add_output_0"
-      ]
+    "version": 1,
+    "model_type": "zipvoice_distill",
+    "encoder": {
+        "name": "encoder_core",
+        "file": "encoder.axmodel",
+        "inputs": [
+            "cat_tokens"
+        ],
+        "outputs": [
+            "encoded"
+        ]
     },
-    {
-      "name": "fm_decoder_part3_1",
-      "file": "decoder_part3_1.axmodel",
-      "inputs": [
-        "/fm_decoder/4/0/bypass/Add_output_0",
-        "time_hidden",
-        "padding_mask"
-      ],
-      "outputs": [
-        "/fm_decoder/4/1/bypass/Add_output_0"
-      ]
-    },
-    {
-      "name": "fm_decoder_part3_2",
-      "file": "decoder_part3_2.axmodel",
-      "inputs": [
-        "/fm_decoder/4/1/bypass/Add_output_0",
-        "time_hidden",
-        "padding_mask"
-      ],
-      "outputs": [
-        "/fm_decoder/4/2/bypass/Add_output_0"
-      ]
-    },
-    {
-      "name": "fm_decoder_part3_3",
-      "file": "decoder_part3_3.axmodel",
-      "inputs": [
-        "/fm_decoder/4/2/bypass/Add_output_0",
-        "time_hidden",
-        "padding_mask"
-      ],
-      "outputs": [
-        "v"
-      ]
-    }
-  ]
-}
+    "decoder_parts": [
+        {
+            "name": "fm_decoder_part0",
+            "file": "decoder_part0.axmodel",
+            "inputs": [
+                "t",
+                "x",
+                "text_condition",
+                "speech_condition",
+                "guidance_scale",
+                "padding_mask"
+            ],
+            "outputs": [
+                "decoder_hidden_p0",
+                "time_hidden"
+            ]
+        },
+        {
+            "name": "fm_decoder_part1",
+            "file": "decoder_part1.axmodel",
+            "inputs": [
+                "decoder_hidden_p0",
+                "time_hidden",
+                "padding_mask"
+            ],
+            "outputs": [
+                "decoder_hidden_p1"
+            ]
+        },
+        {
+            "name": "fm_decoder_part2",
+            "file": "decoder_part2.axmodel",
+            "inputs": [
+                "decoder_hidden_p1",
+                "time_hidden",
+                "padding_mask"
+            ],
+            "outputs": [
+                "decoder_hidden_p2"
+            ]
+        },
+        {
+            "name": "fm_decoder_part3",
+            "file": "decoder_part3.axmodel",
+            "inputs": [
+                "decoder_hidden_p2",
+                "time_hidden",
+                "padding_mask"
+            ],
+            "outputs": [
+                "v"
+            ]
+        }
+    ]
+}

models/zipvoice_distill_ax630C/decoder_part0.axmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2113f394363a575904669017de685b4d8d73693736a285a4ce34782eba1a4531
-size 38600656
+oid sha256:1f0bd9974d94f349c0001d624d3c2c7140295abbe9c12c0db3556d5759d845a8
+size 42595736

models/zipvoice_distill_ax630C/decoder_part1.axmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6afd89c302f971b10677f27093bf60e7cb8f81ec191cde5c43781954976142be
-size 32510786
+oid sha256:cfe296e53b7b8f3842aea1bbc60e48f946c0b5f0ae57c639faa4d31d75ff9c5e
+size 31881170

models/zipvoice_distill_ax630C/decoder_part2.axmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c883008c38b1ee727b8a68989dd4861d368a1cb593f4548e44fc6200dee89e6d
-size 35030020
+oid sha256:e95d65c96ba38c7a3cee6ecc558567a4d79e1803ba427749442690afc3b99da9
+size 35101628

models/zipvoice_distill_ax630C/decoder_part3_0.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0b4d6a68bbc3286ed14bbe1d7f783bcc8cd559fe709151a99079d6558c1ce7d9
-size 10038370

models/zipvoice_distill_ax630C/decoder_part3_1.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:744d3121d74670fec1f29f9a7ce5d158cb215e9b822464ad123fe8daeb8247f1
-size 10038193

models/zipvoice_distill_ax630C/decoder_part3_2.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:719c53fb3e4e0949b893c25814f705d33b12cd442fb84434d00dc3867c846ac1
-size 10038201

models/zipvoice_distill_ax630C/decoder_part3_3.axmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7c44654f65f1398e611b02d9da2a23e931008ffa9ee6a53ebf027fd0da04bd17
-size 10004459

models/zipvoice_distill_ax630C/encoder.axmodel

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7193cbc302f3063522194b8287e1adfc250b83ab67c38c064bba74bef3eacb94
-size 7659503
+oid sha256:ffbd78ded4a05cf472f78a66e34d9c4dac4c5525f6a7b1d2cc5dbb4d96e7538f
+size 8394155

run_ax630c.sh

@@ -1,51 +0,0 @@
-#!/bin/bash
-# ZipVoice AX630C 板端一键运行
-# 用法: bash run_ax630c.sh [zh|en] [sentence|paragraph]
-
-MODEL="${1:-zh}"
-MODE="${2:-sentence}"
-
-BIN="bin/zipvoice_ax630c"
-MODEL_DIR="./models/zipvoice_distill_ax630C"
-TOKEN="./resources/zipvoice_hf/zipvoice/tokens.txt"
-VOCODER="./models/vocoder/vocos_full_ax630c.axmodel"
-
-cd "$(dirname "$0")"
-
-if [ "$MODEL" = "en" ]; then
-    PWAV="./assets/moss_prompts/en_4_4p5s.wav"
-    PTEXT="This is almost twice the current industry production level per train."
-    DTEXT="This morning, a small train left the station, carrying sleepy passengers toward a bright coastal town."
-    TFILE="./assets/paragraphs/en_scavenger.txt"
-    REPO="--repo-dir ."
-else
-    PWAV="./assets/moss_prompts/zh_1_4p5s.wav"
-    PTEXT="不管怎么样我和汤姆还是要感谢贝尔卡金的援手"
-    DTEXT="今天午后天气很好，我打开窗户，听见远处有人聊天，水杯也轻轻晃了一下。"
-    TFILE="./assets/paragraphs/zh_ginkgo.txt"
-    REPO=""
-fi
-
-OUT="output_ax630c_${MODEL}_${MODE}.wav"
-
-echo "=========================================="
-echo " ZipVoice AX630C"
-echo "=========================================="
-echo " MODEL:  $MODEL_DIR"
-echo " LANG:   $MODEL  MODE: $MODE"
-echo " OUTPUT: $OUT"
-echo "=========================================="
-
-if [ "$MODE" = "paragraph" ]; then
-    exec $BIN \
-        --model-dir "$MODEL_DIR" --token-file "$TOKEN" \
-        --prompt-wav "$PWAV" --prompt-text "$PTEXT" \
-        --text-file "$TFILE" \
-        --vocoder-model "$VOCODER" --output-wav "$OUT" --seed 42 $REPO
-else
-    exec $BIN \
-        --model-dir "$MODEL_DIR" --token-file "$TOKEN" \
-        --prompt-wav "$PWAV" --prompt-text "$PTEXT" \
-        --text "$DTEXT" \
-        --vocoder-model "$VOCODER" --output-wav "$OUT" --seed 42 $REPO
-fi

run_ax650.sh

@@ -1,57 +0,0 @@
-#!/bin/bash
-# ZipVoice AXERA 板端一键运行
-# 用法: bash run_ax650.sh [standard|distill] [zh|en] [sentence|paragraph]
-
-MODEL="${1:-distill}"
-LANG="${2:-zh}"
-MODE="${3:-sentence}"
-
-BIN="bin/zipvoice_ax650"
-TOKEN="./resources/zipvoice_hf/zipvoice/tokens.txt"
-VOCODER="./models/vocoder/vocos_full.axmodel"
-
-cd "$(dirname "$0")"
-
-if [ "$MODEL" = "distill" ]; then
-    MODEL_DIR="./models/zipvoice_distill_ax650"
-else
-    MODEL_DIR="./models/zipvoice_ax650"
-fi
-
-if [ "$LANG" = "en" ]; then
-    PWAV="./assets/moss_prompts/en_4_4p5s.wav"
-    PTEXT="This is almost twice the current industry production level per train."
-    DTEXT="This morning, a small train left the station, carrying sleepy passengers toward a bright coastal town."
-    TFILE="./assets/paragraphs/en_scavenger.txt"
-    REPO="--repo-dir ."
-else
-    PWAV="./assets/moss_prompts/zh_1_4p5s.wav"
-    PTEXT="不管怎么样我和汤姆还是要感谢贝尔卡金的援手"
-    DTEXT="今天午后天气很好，我打开窗户，听见远处有人聊天，水杯也轻轻晃了一下。"
-    TFILE="./assets/paragraphs/zh_ginkgo.txt"
-    REPO=""
-fi
-
-OUT="output_${LANG}_${MODEL}_${MODE}.wav"
-
-echo "=========================================="
-echo " ZipVoice AX650"
-echo "=========================================="
-echo " MODEL:  $MODEL_DIR"
-echo " LANG:   $LANG  MODE: $MODE"
-echo " OUTPUT: $OUT"
-echo "=========================================="
-
-if [ "$MODE" = "paragraph" ]; then
-    exec $BIN \
-        --model-dir "$MODEL_DIR" --token-file "$TOKEN" \
-        --prompt-wav "$PWAV" --prompt-text "$PTEXT" \
-        --text-file "$TFILE" \
-        --vocoder-model "$VOCODER" --output-wav "$OUT" --seed 42 $REPO
-else
-    exec $BIN \
-        --model-dir "$MODEL_DIR" --token-file "$TOKEN" \
-        --prompt-wav "$PWAV" --prompt-text "$PTEXT" \
-        --text "$DTEXT" \
-        --vocoder-model "$VOCODER" --output-wav "$OUT" --seed 42 $REPO
-fi

run_axcl.sh

@@ -1,57 +0,0 @@
-#!/bin/bash
-# ZipVoice AXCL 算力卡一键运行
-# 用法: bash run_axcl.sh [standard|distill] [zh|en] [sentence|paragraph]
-
-MODEL="${1:-distill}"
-LANG="${2:-zh}"
-MODE="${3:-sentence}"
-
-BIN="bin/zipvoice_axcl"
-TOKEN="./resources/zipvoice_hf/zipvoice/tokens.txt"
-VOCODER="./models/vocoder/vocos_full.axmodel"
-
-cd "$(dirname "$0")"
-
-if [ "$MODEL" = "distill" ]; then
-    MODEL_DIR="./models/zipvoice_distill_ax650"
-else
-    MODEL_DIR="./models/zipvoice_ax650"
-fi
-
-if [ "$LANG" = "en" ]; then
-    PWAV="./assets/moss_prompts/en_4_4p5s.wav"
-    PTEXT="This is almost twice the current industry production level per train."
-    DTEXT="This morning, a small train left the station, carrying sleepy passengers toward a bright coastal town."
-    TFILE="./assets/paragraphs/en_scavenger.txt"
-    REPO="--repo-dir ."
-else
-    PWAV="./assets/moss_prompts/zh_1_4p5s.wav"
-    PTEXT="不管怎么样我和汤姆还是要感谢贝尔卡金的援手"
-    DTEXT="今天午后天气很好，我打开窗户，听见远处有人聊天，水杯也轻轻晃了一下。"
-    TFILE="./assets/paragraphs/zh_ginkgo.txt"
-    REPO=""
-fi
-
-OUT="output_${LANG}_${MODEL}_${MODE}.wav"
-
-echo "=========================================="
-echo " ZipVoice AXCL"
-echo "=========================================="
-echo " MODEL:  $MODEL_DIR"
-echo " LANG:   $LANG  MODE: $MODE"
-echo " OUTPUT: $OUT"
-echo "=========================================="
-
-if [ "$MODE" = "paragraph" ]; then
-    exec $BIN \
-        --model-dir "$MODEL_DIR" --token-file "$TOKEN" \
-        --prompt-wav "$PWAV" --prompt-text "$PTEXT" \
-        --text-file "$TFILE" \
-        --vocoder-model "$VOCODER" --output-wav "$OUT" --seed 42 $REPO
-else
-    exec $BIN \
-        --model-dir "$MODEL_DIR" --token-file "$TOKEN" \
-        --prompt-wav "$PWAV" --prompt-text "$PTEXT" \
-        --text "$DTEXT" \
-        --vocoder-model "$VOCODER" --output-wav "$OUT" --seed 42 $REPO
-fi

scripts/common_infer.py

@@ -66,74 +66,3 @@ def vocoder_decode_loaded(
     if prompt_rms < target_rms:
         wav = wav * prompt_rms / target_rms
     return wav.squeeze().cpu().numpy()
-
-
-def load_axmodel_vocoder(model_path: str):
-    """Load vocoder axmodel for inference."""
-    import axengine as axe
-
-    session = axe.InferenceSession(model_path)
-    return session
-
-
-def axmodel_vocoder_decode(session, features: np.ndarray, feat_scale: float,
-                            target_rms: float, prompt_rms: float) -> np.ndarray:
-    """Decode mel features to audio using axmodel vocoder + IRFFT."""
-    import math
-
-    # features shape: (1, T, 100) or (T, 100), squeeze batch dim
-    features = np.squeeze(features)
-    if features.ndim == 2:
-        features = features.T  # (T, 100) → (100, T)
-    T = features.shape[1]
-    n_mels = features.shape[0]
-    n_fft, hop = 1024, 256
-    T_model = 620
-
-    # Undo feat_scale and pad to [1, n_mels, T_model]
-    inv_scale = 1.0 / feat_scale if feat_scale != 0 else 1.0
-    mel_input = np.zeros((1, n_mels, T_model), dtype=np.float32)
-    t_actual = min(T, T_model)
-    for t in range(t_actual):
-        mel_input[0, :, t] = features[:, t] * inv_scale
-
-    # Run axmodel: mel → (real, imag)
-    real, imag = session.run(None, {"mel": mel_input})
-    real, imag = real.squeeze(0), imag.squeeze(0)  # [T_model, n_freqs]
-
-    # IRFFT + overlap-add (matching C++ vocoder)
-    n_freqs = n_fft // 2 + 1
-    window = 0.5 * (1.0 - np.cos(2.0 * math.pi * np.arange(n_fft) / (n_fft - 1)))
-    window_sq = window ** 2
-    out_len = (T - 1) * hop + n_fft
-    audio = np.zeros(out_len, dtype=np.float32)
-    envelope = np.zeros(out_len, dtype=np.float32)
-
-    for t_idx in range(T):
-        spec = np.zeros(n_fft, dtype=np.complex64)
-        spec[0] = real[t_idx, 0]
-        for k in range(1, n_freqs - 1):
-            spec[k] = real[t_idx, k] + 1j * imag[t_idx, k]
-            spec[n_fft - k] = real[t_idx, k] - 1j * imag[t_idx, k]
-        spec[n_freqs - 1] = real[t_idx, n_freqs - 1]
-        ifft_out = np.fft.irfft(spec, n=n_fft).real
-
-        pos = t_idx * hop
-        for n in range(n_fft):
-            p = pos + n
-            if p < out_len:
-                audio[p] += ifft_out[n] * window[n]
-                envelope[p] += window_sq[n]
-
-    audio /= np.maximum(envelope, 1e-10)
-    pad = n_fft // 2
-    audio = audio[pad:out_len - pad].astype(np.float32)
-
-    # RMS normalize (numpy version)
-    rms = np.sqrt(np.mean(audio ** 2))
-    if rms < target_rms and rms > 1e-10:
-        audio = audio * (target_rms / rms)
-    if prompt_rms < target_rms:
-        audio = audio * (prompt_rms / target_rms)
-    return audio
-

scripts/zipvoice_decoder4_runtime.py

@@ -276,7 +276,6 @@ class Decoder4ZipVoiceBoardRuntime:
         encoded = self.run_encoder(cat_tokens)
         t_enc = time.perf_counter() - t_start
         logging.debug("  encoder: %.3f s (output shape=%s)", t_enc, encoded.shape)
-        # DEBUG: dump intermediates
 
         t_start = time.perf_counter()
         text_condition, features_len = self.duration_expand(
@@ -288,7 +287,6 @@ class Decoder4ZipVoiceBoardRuntime:
         )
         t_dur = time.perf_counter() - t_start
         logging.debug("  duration_expand: %.3f s (features_len=%d)", t_dur, features_len)
-        # DEBUG: dump
 
         seq_len = self.decoder_seq_len or self.max_feat_len
         if features_len > seq_len:

scripts/zipvoice_decoder4_runtime_encoder_onnx.py

@@ -1,87 +0,0 @@
-#!/usr/bin/env python3
-
-from __future__ import annotations
-
-import logging
-from pathlib import Path
-from typing import Dict, List
-
-import numpy as np
-import onnxruntime as ort
-
-from scripts.zipvoice_decoder4_runtime import Decoder4ZipVoiceBoardRuntime
-from scripts.zipvoice_runtime import AxeSession
-
-
-class _OrtSession:
-    """ONNX Runtime wrapper with the AxeSession interface used by split runtime."""
-
-    _TYPE_TO_DTYPE = {
-        "tensor(float)": np.float32,
-        "tensor(float32)": np.float32,
-        "tensor(double)": np.float64,
-        "tensor(int32)": np.int32,
-        "tensor(int64)": np.int64,
-        "tensor(uint8)": np.uint8,
-        "tensor(bool)": np.bool_,
-    }
-
-    def __init__(self, model_path: str | Path):
-        self.path = Path(model_path)
-        if not self.path.exists():
-            raise FileNotFoundError(f"ONNX model not found: {self.path}")
-        self._session = ort.InferenceSession(
-            str(self.path),
-            providers=["CPUExecutionProvider"],
-        )
-        self._inputs = self._session.get_inputs()
-        self._outputs = self._session.get_outputs()
-
-    @property
-    def input_names(self) -> List[str]:
-        return [item.name for item in self._inputs]
-
-    @property
-    def output_names(self) -> List[str]:
-        return [item.name for item in self._outputs]
-
-    def _coerce_one(self, name: str, value: np.ndarray) -> np.ndarray:
-        info = next((item for item in self._inputs if item.name == name), None)
-        array = np.asarray(value)
-        if info is None:
-            return np.ascontiguousarray(array)
-
-        dtype = self._TYPE_TO_DTYPE.get(str(info.type).lower())
-        if dtype is not None:
-            array = array.astype(dtype, copy=False)
-
-        if list(info.shape) == [] and array.shape == (1,):
-            array = array.reshape(())
-
-        return np.ascontiguousarray(array)
-
-    def run(self, feed_dict: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
-        feed = {name: self._coerce_one(name, value) for name, value in feed_dict.items()}
-        outputs = self._session.run(None, feed)
-        return {name: value for name, value in zip(self.output_names, outputs)}
-
-
-class Decoder4ZipVoiceBoardRuntimeEncoderOnnx(Decoder4ZipVoiceBoardRuntime):
-    """Runs encoder with ONNX Runtime, all decoder parts with axmodel."""
-
-    def _load_models(self) -> None:
-        self.sessions = {}
-
-        encoder_name = self.encoder_info["name"]
-        encoder_path = self.models_dir / "encoder_core.onnx"
-        logging.info("encoder 使用 ONNX Runtime: %s", encoder_path)
-        self.sessions[encoder_name] = _OrtSession(encoder_path)
-
-        for info in self.decoder_parts:
-            name = info["name"]
-            path = self.models_dir / info["file"]
-            logging.debug("Loading %s from %s", name, path)
-            self.sessions[name] = AxeSession(path)
-
-        self.decoder_label = "decoder4(encoder_onnx+part0-3_axmodel)"
-        logging.debug("Loaded encoder ONNX + %d decoder axmodels", len(self.decoder_parts))

scripts/zipvoice_decoder4_runtime_part0_onnx.py

@@ -1,94 +0,0 @@
-#!/usr/bin/env python3
-
-from __future__ import annotations
-
-import logging
-from pathlib import Path
-from typing import Dict, List
-
-import numpy as np
-import onnxruntime as ort
-
-from scripts.zipvoice_decoder4_runtime import Decoder4ZipVoiceBoardRuntime
-from scripts.zipvoice_runtime import AxeSession
-
-
-class _OrtSession:
-    """ONNX Runtime wrapper with the small AxeSession interface used by the board runtime."""
-
-    _TYPE_TO_DTYPE = {
-        "tensor(float)": np.float32,
-        "tensor(float32)": np.float32,
-        "tensor(double)": np.float64,
-        "tensor(int32)": np.int32,
-        "tensor(int64)": np.int64,
-        "tensor(uint8)": np.uint8,
-        "tensor(bool)": np.bool_,
-    }
-
-    def __init__(self, model_path: str | Path):
-        self.path = Path(model_path)
-        if not self.path.exists():
-            raise FileNotFoundError(f"ONNX model not found: {self.path}")
-        self._session = ort.InferenceSession(
-            str(self.path),
-            providers=["CPUExecutionProvider"],
-        )
-        self._inputs = self._session.get_inputs()
-        self._outputs = self._session.get_outputs()
-
-    @property
-    def input_names(self) -> List[str]:
-        return [item.name for item in self._inputs]
-
-    @property
-    def output_names(self) -> List[str]:
-        return [item.name for item in self._outputs]
-
-    def _coerce_one(self, name: str, value: np.ndarray) -> np.ndarray:
-        info = next((item for item in self._inputs if item.name == name), None)
-        array = np.asarray(value)
-        if info is None:
-            return np.ascontiguousarray(array)
-
-        dtype = self._TYPE_TO_DTYPE.get(str(info.type).lower())
-        if dtype is not None:
-            array = array.astype(dtype, copy=False)
-
-        # The exported part0 ONNX keeps t/guidance_scale as scalar inputs ([]),
-        # while the axmodel path feeds them as shape [1]. Normalize only for ONNX.
-        if list(info.shape) == [] and array.shape == (1,):
-            array = array.reshape(())
-
-        return np.ascontiguousarray(array)
-
-    def run(self, feed_dict: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
-        feed = {name: self._coerce_one(name, value) for name, value in feed_dict.items()}
-        outputs = self._session.run(None, feed)
-        return {name: value for name, value in zip(self.output_names, outputs)}
-
-
-class Decoder4ZipVoiceBoardRuntimePart0Onnx(Decoder4ZipVoiceBoardRuntime):
-    """Runs decoder part0 with ONNX Runtime, all other split models with axmodel."""
-
-    def _load_models(self) -> None:
-        self.sessions = {}
-
-        encoder_name = self.encoder_info["name"]
-        encoder_path = self.models_dir / self.encoder_info["file"]
-        logging.debug("Loading %s from %s", encoder_name, encoder_path)
-        self.sessions[encoder_name] = AxeSession(encoder_path)
-
-        for index, info in enumerate(self.decoder_parts):
-            name = info["name"]
-            if index == 0:
-                path = self.models_dir / "fm_decoder_part0.onnx"
-                logging.info("part0 使用 ONNX Runtime: %s", path)
-                self.sessions[name] = _OrtSession(path)
-            else:
-                path = self.models_dir / info["file"]
-                logging.debug("Loading %s from %s", name, path)
-                self.sessions[name] = AxeSession(path)
-
-        self.decoder_label = "decoder4(part0_onnx+part1-3_axmodel)"
-        logging.debug("Loaded encoder axmodel + part0 ONNX + %d decoder axmodels", len(self.decoder_parts) - 1)

scripts/zipvoice_decoder4_runtime_part3_onnx.py

@@ -1,93 +0,0 @@
-#!/usr/bin/env python3
-
-from __future__ import annotations
-
-import logging
-from pathlib import Path
-from typing import Dict, List
-
-import numpy as np
-import onnxruntime as ort
-
-from scripts.zipvoice_decoder4_runtime import Decoder4ZipVoiceBoardRuntime
-from scripts.zipvoice_runtime import AxeSession
-
-
-class _OrtSession:
-    """ONNX Runtime wrapper with the AxeSession interface used by split runtime."""
-
-    _TYPE_TO_DTYPE = {
-        "tensor(float)": np.float32,
-        "tensor(float32)": np.float32,
-        "tensor(double)": np.float64,
-        "tensor(int32)": np.int32,
-        "tensor(int64)": np.int64,
-        "tensor(uint8)": np.uint8,
-        "tensor(bool)": np.bool_,
-    }
-
-    def __init__(self, model_path: str | Path):
-        self.path = Path(model_path)
-        if not self.path.exists():
-            raise FileNotFoundError(f"ONNX model not found: {self.path}")
-        self._session = ort.InferenceSession(
-            str(self.path),
-            providers=["CPUExecutionProvider"],
-        )
-        self._inputs = self._session.get_inputs()
-        self._outputs = self._session.get_outputs()
-
-    @property
-    def input_names(self) -> List[str]:
-        return [item.name for item in self._inputs]
-
-    @property
-    def output_names(self) -> List[str]:
-        return [item.name for item in self._outputs]
-
-    def _coerce_one(self, name: str, value: np.ndarray) -> np.ndarray:
-        info = next((item for item in self._inputs if item.name == name), None)
-        array = np.asarray(value)
-        if info is None:
-            return np.ascontiguousarray(array)
-
-        dtype = self._TYPE_TO_DTYPE.get(str(info.type).lower())
-        if dtype is not None:
-            array = array.astype(dtype, copy=False)
-
-        if list(info.shape) == [] and array.shape == (1,):
-            array = array.reshape(())
-
-        return np.ascontiguousarray(array)
-
-    def run(self, feed_dict: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
-        feed = {name: self._coerce_one(name, value) for name, value in feed_dict.items()}
-        outputs = self._session.run(None, feed)
-        return {name: value for name, value in zip(self.output_names, outputs)}
-
-
-class Decoder4ZipVoiceBoardRuntimePart3Onnx(Decoder4ZipVoiceBoardRuntime):
-    """Runs decoder part3 with ONNX Runtime, encoder and part0-2 with axmodel."""
-
-    def _load_models(self) -> None:
-        self.sessions = {}
-
-        encoder_name = self.encoder_info["name"]
-        encoder_path = self.models_dir / self.encoder_info["file"]
-        logging.debug("Loading %s from %s", encoder_name, encoder_path)
-        self.sessions[encoder_name] = AxeSession(encoder_path)
-
-        last_index = len(self.decoder_parts) - 1
-        for index, info in enumerate(self.decoder_parts):
-            name = info["name"]
-            if index == last_index:
-                path = self.models_dir / "fm_decoder_part3.onnx"
-                logging.info("part3 使用 ONNX Runtime: %s", path)
-                self.sessions[name] = _OrtSession(path)
-            else:
-                path = self.models_dir / info["file"]
-                logging.debug("Loading %s from %s", name, path)
-                self.sessions[name] = AxeSession(path)
-
-        self.decoder_label = "decoder4(part0-2_axmodel+part3_onnx)"
-        logging.debug("Loaded encoder axmodel + %d decoder axmodels + part3 ONNX", last_index)