一句话：VoxCPM 2 是 OpenBMB 开源的语音模型，代码和权重都是 Apache-2.0，可以商用。VoiceStudio 是一个开源的本地配音桌面应用，可以把 VoxCPM 2 装成它的一个引擎，提供声音库、转写和本地 API。两者配合，在自己电脑上就能做出接近真人的中文旁白，不按字数付费，稿子和录音也不用上传。
本文要点
- 先科普：配音为什么难；"预设音色、声音设计、声音克隆"三种玩法各适合什么；两个项目截至 2026 年 10 月 30 日的状态。
- 再实战：两条路线。
- 路线 A：VoiceStudio 图形界面加 VoxCPM 2 引擎，零代码。
- 路线 B：Mac 上直接用 mlx-audio 跑 VoxCPM 2，适合批量生产，也是我们一直在用的方式。
- 最值钱的部分：参考音频怎么洗、长稿怎么分段、多音字和英文词怎么处理、怎么质检。
- 弯路清单：一个月里踩过的 15 个坑，每个都有症状、原因和解法。
- 实测数据：Apple M3 Max 上的速度和内存、参考音频清理前后的对比、多音字的小样本实验。
为什么写这篇
做视频的朋友，几乎都在配音上吃过苦头：
- 自己录：要安静的房间，要好麦克风，录完还得剪口误。改一个字，整段重录。
- 用免费的在线 TTS：一听就是机器，语调平，停顿怪，观众三秒就划走。
- 花钱买会员、按字数付费：贵，好听的音色还撞衫。你的稿子还要上传到平台，音色也不属于你。
- 找人配：更贵，周期更长，改稿还得再等。
我们在做开源项目 sweety-handdraw-video（AI 手绘口播视频）的过程中，几乎把这些坑全踩了一遍。后来定型的方案就是 VoiceStudio + VoxCPM 2。
手绘项目 2.0 发布时，我们提了一句"高度拟人化的配音靠的是这个组合"，之后很多人来问怎么做。
准确地说，这两个在我们这里一直是分开跑的：
- 先用 VoiceStudio：当时配的是 Qwen3-TTS 引擎，手绘项目 1.0 的配音和语音转写都靠它。
- VoxCPM 2 后来加入：在 Mac 上用 mlx-audio 单独跑，成了主力旁白，造就了独一无二的逼真配音。
现在 VoiceStudio 已经能把 VoxCPM 2 装成自己的一个引擎，下面我会把这两条线都讲一遍。
第一部分：科普——先把概念理清，少走一半弯路
1. AI 配音的三种玩法
我们最后的做法是两种组合起来：先用"声音设计"找到喜欢的声音，挑一条最好的作参考，以后全部用"克隆"。这样既不用找真人授权，音色又稳定。后面会细讲。
2. VoxCPM 2：现在最值得用的开源中文语音模型之一
VoxCPM 是 OpenBMB 开源的语音生成模型。VoxCPM 2 在 2026 年 4 月 6 日发布 2.0.0。截至本文，最新的 Python 包是 2.0.3，2026 年 5 月 11 日发布，加了 --device 参数，流式解码也更快了。
它最让我们惊喜的是三种生成模式，也是后面所有操作的基础：
新手注意：高保真克隆模式下，括号里的语气指令会被忽略。想要"同一个声音，换种语气"，应该用可控克隆，而不是在高保真模式里拼命写指令。
在 Seed-TTS-eval 上的成绩如下（这是 OpenBMB 自己发布的表，未经第三方核验）：
来源：VoxCPM README · Performance。
3. VoiceStudio
VoiceStudio 是开源的本地配音桌面应用，自称"开源、全本地的 ElevenLabs 替代品"，有中文 README。
它的定位是工作台加引擎管理器，自己不绑定某一个模型：
- 多引擎：默认引擎是 OmniVoice，还可以装 VoxCPM 2、CosyVoice 3、IndexTTS、MOSS-TTS、GPT-SoVITS、MLX-Audio 等。
- 声音库：参考录音、逐字稿、种子都存成"声音档案"，下次直接选。
- 配套工具：语音转写（Whisper）、人声分离、视频配音、有声书和长文（支持 [pause 0.5s] 停顿标记）、发音词典。
- 本地 API：默认端口 3900，有 OpenAI 兼容的 /v1/audio/speech 和 /v1/audio/transcriptions，还有 MCP 接口，可以接 Claude Code、Codex 这类 Agent。
最新版本是 0.5.7，2026 年 10 月 26 日发布，有几处要注意：
- 许可确认：没有明确商用许可的模型，包括默认引擎 OmniVoice，要先在应用里接受一次许可，否则 API 和 MCP 调用会失败。
- 新安装包和界面：新增纯 CPU 版本，界面重新设计。
- 付费功能：远程和 GPU 共享这类功能移到了付费的 Pro 版。
4. 为什么是 VoiceStudio + VoxCPM 2的组合？
一个很多人没注意到的事实：VoiceStudio 应用本身是 AGPL-3.0，但它的默认引擎 OmniVoice 的模型权重是 CC-BY-NC（非商业）（模型卡）。
如果你做的是带货、接广告、知识付费这类商业内容，默认引擎在许可上就不合适。VoxCPM 2 的权重是 Apache-2.0，商用没有这个顾虑，中文和方言能力也强。
所以分工是：
- VoiceStudio 负责好用：界面、声音库、转写、配音工作台、本地 API。
- VoxCPM 2 负责好听和能商用。
提醒：模型许可允许商用，不等于你可以克隆任何人的声音。声音授权是另一回事，见文末"合规"部分。
第二部分：我们的心路历程
在讲操作之前，先讲我们怎么一步步走到现在的方案。
阶段一：工具大杂烩
最早每个项目各用各的：Edge TTS、Piper、Kokoro、Qwen3-TTS 的两个版本、Whisper.cpp、单独装的 MLX Audio……每个都能用，没有一个好用：
- 有的中文生硬；
- 有的只有几个预设音色；
- 有的能克隆，但需要单独的环境。
2026 年 10 月初我们下决心清理，卸掉这些独立工具和重复模型，腾出了约 13 GB 磁盘，配音统一收口到 VoiceStudio。
教训：先选一个统一入口。 工具越多，环境越乱，你会把时间花在这次用的是哪个上。
阶段二：情绪调不出来
收口后，我们先用 VoiceStudio 加 Qwen3-TTS 的克隆模式给绘本配音：妈妈和孩子两个角色，各自先用"声音设计"生成参考，再克隆。
问题来了：妈妈的台词需要温柔、着急、开心几种情绪，可不管指令写得多细，出来都一个调。
我们翻了源码才发现：当时 VoiceStudio 里跑的 Qwen3-TTS 克隆分支是 mlx-audio 的实现，它接收了情绪指令，却根本没有把指令传给模型。
教训：情绪不出来时，先确认你用的模式和引擎到底支不支持情绪指令。别急着改提示词。 这也是我们后来转向 VoxCPM 2 的原因之一：它的"可控克隆"明确支持用指令调语气。
阶段三：VoxCPM 2 上手，音色惊艳
我们在 Mac 上用 mlx-audio 跑 VoxCPM 2 的 8-bit 版本。我们没有把它装进 VoiceStudio，而是单独建了一个 Python 环境直接跑。第一次克隆出来的男声，音色质感比之前所有方案都好。我们当时偷懒，把整篇稿子一次性生成，47 秒一口气读完。
成片听下来，背景里一直有若有若无的嗡声。我们一开始以为是模型的问题，还试着调参数。
阶段四：调优
做了频谱分析才找到根因：我们那段 7 秒的参考录音，是从一段带背景音乐的旁白里直接截的。
- 配乐残留：频谱上有 106、593、891 Hz 三条持续的"线"。
- 停顿不安静：停顿处的音量只比说话低 17.8 dB，正常的干声应该低 25 dB 以上。
- 轻微削波：录音有轻微的破音。
模型把配乐当成了"这个人声音的一部分"，一起克隆了进来，而且越长的段落放得越大。
修法：回到原始录音，挑一段 14.9 秒、在句末结束的完整段落，先用 HTDemucs 做人声分离，再用 DeepFilterNet3 降噪。结果持续线谱消失，停顿比语音低 33.5 dB，语音识别的字错误率（CER）0.035。
上图左边是清理前，能看到几条贯穿全程的横线，那就是配乐，停顿处也一直"不安静"。右边是清理后，横线没了，停顿深深地沉下去。
另一款女声也经历了同样的清理。重新生成后，我们的反馈原话是："音色接近，杂音也消失了"，全篇试听后"全篇也通过，作为主力女声"。
参考音频的质量，决定了克隆的上限。
阶段五：长稿的玄学
接下来做的片子越来越长，从一两分钟到十一分钟。新问题一个接一个，后面会逐条列，这里先挑几个印象深的：
- 整篇一次生成：越读越快，偶尔出现嗡声，甚至停不下来。官方也承认长文本不稳定，建议分段。
- 句尾被吃字：有一句"多试几遍"，成片里"遍"没了。原因是我们裁句尾静音的算法，把"遍"字开口前那几十毫秒的闭塞（发 b 音前嘴唇闭住的那一下）当成了句子结束。
- 英文单词听不懂：bug、Thorsten 这类词，同一句话换了 8 个种子，一次都没读对。
- 机器质检误报：男声念一长串列举时，音高会在 120 Hz 附近稳住好几秒，被我们的嗡声检测当成了"杂音"。同一句话 8 个种子，7 个被误判。
每个问题我们都把规则写进了流程里，后面几节就是这些规则。
阶段六：克隆自己的声音
我们也试过克隆自己。从一段会议录音里截出本人主讲的 9 秒，降噪后门禁通过，识别字错误率为 0，技术上无可挑剔。
但放进正式的知识解说片，自己听着就是不对味：日常说话的声音，不等于适合做旁白的声音。最后我们的分工是：
- 个人口播、自媒体：用自己的声音。
- 正式解说片：用一款更有"播音感"的男声。
教训：技术合格不等于适合。 选声音时，用你要发布的那类内容试听，别只用一句"你好，欢迎来到我的频道"。
阶段七：多音字
最近一次给一段家庭羽毛球视频配解说，21 句旁白里，ASR 质检揪出了这些读错的：
我们的处理：
- 换种子重生成；
- 实在不行就改写说法，比如把"夜场羽毛球"开场改成别的说法；
- 最后只剩"十几拍"一句：ASR 仍听写成"实际拍"。两者只差声调，ASR 判声调本身就不可靠，所以留给人耳确认，其余全部过关。
后面"多音字"一节，有我们专门做的小实验。
第三部分：实战路线 A —— VoiceStudio + VoxCPM 2（零代码）
适合绝大多数人：不想碰命令行，想要一个"打开就能用"的配音软件。
第 0 步：先确认你的电脑够不够
磁盘：VoxCPM 2 引擎有自己独立的 Python 环境。0.5.6 安装时会先检查是否有约 10 GB 可用空间，首次生成还会下载约 5 GB 模型。建议至少留出 20 GB 空间。
第 1 步：安装 VoiceStudio
macOS 或 Linux，打开"终端"，粘贴：
curl -fsSL https://voicestudio.sh/install | sh
Windows，打开 PowerShell：
irm https://voicestudio.sh/install.ps1 | iex
也可以去 Releases 页面下载安装包：Mac 用 DMG，Windows 用 .exe。Windows 安装包没有签名，SmartScreen 会弹警告，确认来源是官方 GitHub 再继续。
不会用终端？VoiceStudio 官方提供了"让 Agent 帮你装"的提示词。把下面这段贴给 Claude Code 或 Codex：
Install the VoiceStudio Electron app on this device and verify it works, following
https://github.com/debpalash/VoiceStudio/blob/main/docs/install/agent.md
第 2 步：国内网络先配镜像
首次使用要从 Hugging Face 下载模型，国内网络常常很慢或连不上。VoiceStudio 的设置里有 Hugging Face 镜像选项，可以切到 hf-mirror.com，它自带连通性测试。先把这个配好，再装引擎，能省掉一大半"下载卡住"的问题。
第 3 步：安装 VoxCPM 2 引擎
按官方文档：模型目录（Model Catalogue）→ VoxCPM2 → 安装（Install）。它会装进一个独立的 Python 环境，不影响其他引擎。
- 安装好后，第一次生成时才会下载模型权重，第一次会慢一些，耐心等。
- 0.5.7 起，默认引擎 OmniVoice 等非商用模型要先接受许可才能调用。VoxCPM 2 是 Apache-2.0，不受这个限制。
Mac 用户注意：VoiceStudio 里还有一个"MLX-Audio"引擎，也列出了 VoxCPM 2。但按 VoiceStudio 自己的文档，经这个引擎调用时参考音频会被忽略，输出也降到 24 kHz。要做克隆，请选上面那个专门的 VoxCPM2 引擎。来源：VoiceStudio mlx-audio 文档。
第 4 步：准备参考音频（最重要，单独成节）
请先跳到下面的"第五部分：参考音频——决定成败的一小时"把参考录音准备好，再回来继续。
第 5 步：建一个"声音档案"
在"声音克隆"页面添加参考录音，填写逐字稿，也就是参考录音里说的每一个字，一字不差，标点照录。保存成一个声音档案，起个好记的名字，比如"栏目男声-v1"。
VoiceStudio 处理 VoxCPM 2 参考音频的规则，弄清楚能少踩坑：
- 它会自动裁掉参考首尾的静音，然后最多保留前 30 秒；
- 如果因为超过 30 秒被截断，它会丢掉逐字稿，退回普通克隆，因为截断后的音频和逐字稿对不上了。
所以：参考音频控制在 8–15 秒，不要超过 30 秒。
来源：VoiceStudio VoxCPM2 引擎文档，以及 0.5.6 源码。
第 6 步：三种模式在 VoiceStudio 里怎么"触发"
VoiceStudio 不让你直接选模式，而是根据你填了什么自动决定（我们读了 0.5.6 的源码确认）：
参数建议，都是官方默认，先别动：
- 引导强度（guidance）2.0：调低更自然但可能跑稿，调高更贴稿但难句容易出杂音。
- 步数（steps）10：官方推荐 4–30，越大细节越好、越慢。我们的主力男声用 16。
第 7 步：生成、试听、对稿
生成后，用 VoiceStudio 自带的转写功能把音频转回文字，和稿子对一遍：
- 有没有漏字、多字？
- 结尾的最后一个字在不在？
- 人名、数字、多音字对不对？
不对就换个种子重新生成这一段。注意，转写只能发现明显问题，最终一定要用耳朵听一遍。
第 8 步：长稿不要一次贴整篇
不管哪个模型，长文本都容易越读越快、出杂音。正确做法是分段生成，规则见第六部分。
VoiceStudio 的"有声书 / 长文"功能支持 [pause 0.5s] 这样的停顿标记，还有发音词典，可以统一纠正某个词的读法，见第七部分。
第 9 步（进阶）：用本地 API 批量生成
VoiceStudio 开着时，本机 3900 端口提供 OpenAI 兼容接口。参数名以 0.5.6 的接口说明为准：
# 列出你的声音档案，拿到 voice 的 ID
curl -s http://127.0.0.1:3900/v1/audio/voices
# 用 VoxCPM 2 引擎、指定声音档案生成一段旁白
curl -s http://127.0.0.1:3900/v1/audio/speech \
-H "Content-Type: application/json" \
-d '{
"model": "voxcpm2",
"voice": "<你的声音档案ID>",
"input": "今天我们用十分钟，讲清楚一个问题。",
"response_format": "wav",
"seed": 2026,
"num_step": 10,
"guidance_scale": 2.0
}' \
-o part-001.wav
- seed 很重要：同一个种子、同样的参数，结果可以复现。哪段不满意就只换那段的种子。
- 格式：response_format 默认是 mp3，做视频建议用 wav。
- 接入 Agent：VoiceStudio 也提供 MCP（http://127.0.0.1:3900/mcp），可以让 Claude Code、Codex 直接调用它配音。
第四部分：实战路线 B —— Mac 上直接跑 VoxCPM 2
适合想批量生产、想把配音接进自己流程的人。我们的手绘视频项目就是这么做的：每段旁白自动生成、自动质检、不过关自动换种子重来，最后拼成完整的旁白母带。
1. 安装
python3 -m venv ~/voxcpm2-mlx
source ~/voxcpm2-mlx/bin/activate
pip install -U mlx-audio soundfile mlx-whisper
# 国内网络可先设镜像
export HF_ENDPOINT=https://hf-mirror.com
mlx-audio（MIT 许可）从 0.4.4 开始支持 VoxCPM 2，截至本文最新是 0.6.3。我们的生产环境是 0.5.8，下面的代码在 0.5.8 上验证过；新版参数名如有变化，以官方说明为准。
权重有三档：
2. 高保真克隆一段旁白
import mlx.core as mx
import numpy as np
import soundfile as sf
from mlx_audio.tts.utils import load
model = load("mlx-community/VoxCPM2-8bit") # 首次会下载约 3.2 GB
ref = "reference.wav" # 洗干净的参考，8–15 秒
ref_text = open("reference.txt", encoding="utf-8").read().strip() # 一字不差的逐字稿
text = "今天我们用十分钟，把这个问题从头讲清楚。"
mx.random.seed(2026) # 固定种子，结果可复现
chunks = list(model.generate(
text=text,
ref_audio=ref, prompt_audio=ref, prompt_text=ref_text, # 三项都给 = 高保真克隆
inference_timesteps=10, cfg_value=2.0,
max_tokens=len(text) * 6 + 10, # 官方"文本长度×6+10"的上限，防止停不下来
))
audio = np.concatenate([np.asarray(c.audio) for c in chunks])
sf.write("part-001.wav", audio, chunks[0].sample_rate) # 48 kHz
另外两种模式的写法：
- 可控克隆：只给 ref_audio，再加 instruct="语速稍慢，语气温和"。
- 声音设计：不给参考，只给 instruct="中年男声，低沉有磁性，沉稳，适合知识解说"。
3. 生成完立刻"对稿"
import mlx_whisper
r = mlx_whisper.transcribe("part-001.wav",
path_or_hf_repo="mlx-community/whisper-large-v3-turbo",
language="zh")
print(r["text"])
把识别结果和原稿逐字比对，算字错误率：
- 超过 0.15，或结尾少了字：这一段作废，换种子重来。
- 自动重试最多 3 次，还不行就改写这句。
4. 实测速度和内存
条件：Apple M3 Max、36 GB 内存，VoxCPM2-8bit，步数 10。
首次加载模型约 19 秒。
换算一下：我们的主力男声语速约 6.6 字/秒，一部 4581 字的稿子，成片旁白约 708 秒（11 分 48 秒）。
别开两个进程并行：我们试过两个生成进程同时跑，总耗时反而更长，还有一次因为内存吃紧被系统杀掉。单进程按顺序跑最稳。
第五部分：参考音频
这是全文最值钱的一节。克隆的质量 80% 由参考决定。
1. 参考音频的标准
2. 声音从哪来？三个合规来源
- 自己录：手机放在安静的衣柜边，离嘴 20 厘米，用平时讲解的语气读一段 15 秒的话。
- 获得授权的他人录音：一定要有明确授权，最好是书面的。
- 用"声音设计"造一个：这是我们最推荐的做法，既不涉及任何真人，又能得到稳定音色：
- 写声音描述，官方建议的写法是"身份 + 质感 + 场景"三段，例如："中年男声，低沉浑厚、有磁性，沉稳有力，适合知识与故事解说"。
- 同一个描述生成 4 次，换不同种子。
- 挑一条最干净、最像你想要的，作为参考。
- 以后全部用克隆模式，音色就固定下来了。
声音描述里只写声音特征，不要写"像某某主播""某平台某音色"，也不要用别家 TTS 的输出当参考。
3. 参考脏了怎么洗
如果你的录音带了配乐或底噪：
图形界面：VoiceStudio 自带人声分离和音频清理功能，把录音先过一遍。
命令行：我们的做法是两步，先分离人声，再降噪。
# 1) 截出 8–15 秒、句末结束的一段（单声道、48 kHz）
ffmpeg -ss 31.6 -to 46.0 -i source.wav -ac 1 -ar 48000 cut.wav
# 2) 人声分离（Demucs，MIT 许可），只保留人声
pip install demucs
demucs --two-stems=vocals cut.wav # 结果在 separated/htdemucs/cut/vocals.wav
# 3) 降噪（DeepFilterNet）
pip install deepfilternet
deepFilter separated/htdemucs/cut/vocals.wav -o cleaned/
洗完怎么判断干不干净？
- 最简单：用 Audacity 之类的软件打开，切到频谱图。如果看到贯穿全程的横线，就是还有配乐或电流声。
- 再戴耳机听停顿处：停顿应该是"死寂"的，而不是"有一层东西"。
- 我们的量化门禁：
- 停顿要比说话低 25 dB 以上，干净的参考能到 −30 dB 以下；
- 不能有持续的单频线；
- 不能削波。
VoxCPM 2 本身也有 denoise 参数，但它只处理参考音频，而且走 16 kHz 管线，官方说会"轻微改变音色"。参考本来就干净的话，保持关闭。
第六部分：长稿生产——分段、停顿、句尾
1. 分段规则
2. 段与段之间留多长停顿
拼接时插入纯静音。下面这组是我们在已验收的成片、以及真人长篇解说上实测的：
3. 句尾怎么处理
模型说完一句后，偶尔会"多出一点声音"，比如一口气、半个杂音。我们的处理：
- 从最后一个字的末尾往后找，第一段持续 30 毫秒、低于 −40 dB 的静音，就是句子真正的结尾。
- 别从字的开头找：b、t、g 这类声母开口前，嘴唇或舌头会闭住几十毫秒，那段静音不是句尾。我们的"多试几遍"丢了"遍"字，就是这个原因。
- 结尾后保留 0.12 秒余音，再用至少 50 毫秒的淡出收尾。
- 裁完再识别一次：确认最后一个字还在。
4. 混音
- 旁白整体响度做到 −15 LUFS 左右（短视频平台常用的范围）。
- 背景音乐在说话时压低。
- 整体峰值不超过 −1.5 dBTP。
第七部分：多音字、英文和数字——中文配音的三块硬骨头
1. 多音字：我们做了个小实验
官方给的方法是在文本里写带声调数字的拼音，比如 {ni3}{hao3}，同时把 normalize 关掉（用法指南）。
社区里还有一个土办法：同音字替换。口播稿里把容易读错的字，换成读音唯一的同音字，字幕仍然用原文。VoiceStudio 的发音词典支持按词替换写法，可以把这件事自动化。
我们在自己的环境里各试了 5 个种子。环境是 mlx-audio 0.5.8、8-bit、高保真克隆，结果用 Whisper 判读：
几点说明：
- 样本很小，只是 5 个种子，结论仅供参考。
- MLX 移植版不一定完整支持官方的拼音标注。用官方 PyTorch 版本的朋友，可以优先试拼音标注。
- ASR 判声调本身不可靠。"夜场 / 夜长"这种只差声调的词，Whisper 在两种写法下都给出过两种结果，我们没法据此判断，所以没列进表里。最后一定要人耳确认。
我们现在的实际做法：
- 写稿时就避开：容易读错的词，能换说法就换。
- 换不掉的，用同音字替换，并在字幕里保留原文。
- 逐段 ASR 对稿，读错就换种子，最多 3 次。
- 仍不行，改写这句。
- 最后人耳通听一遍。
2. 英文词：尽量少用
中文语音模型读中文里夹的英文，经常"中文识别能过，英文耳朵听不懂"。我们实测了主力男声（VoxCPM2-8bit 克隆）：
- bug、Thorsten：同一句话 8 个种子，一次都没听出原词。
- git、Amp：部分种子能听出。
- AI、Agent、token：大多能听出。
所以我们定了规矩：
- 常用词有中文说法就用中文，比如 bug 写成"问题"。
- 人名用通行的中文译名，画面上可以保留英文。
- 必须保留的英文，生成后单独切出来，用英文识别再听一遍，听不出就换种子，最多 6 次。
3. 数字、日期、单位：写成汉字
官方的 normalize 参数可以处理数字读法，但 MLX 移植版没有。最稳的办法是在口播稿里直接写成你想读的样子：
- "2026 年"写成"二零二六年"；
- "1024"先决定读"一零二四"还是"一千零二十四"；
- "2.0.3"写成"二点零点三"。
第八部分：我们踩过的 15 个坑（速查表）
还有一条不算坑，算心得：克隆自己的声音，技术上满分，不一定适合正式内容。 选声音要用真实内容试听。
第九部分：最终效果
- 音色：两款主力声音（一男一女）参考清理后，嗡声、底噪消失。女声重新生成后的试听反馈是"音色接近，杂音也消失了"，整篇通过。
- 稳定：同一个参考加固定种子，结果可复现。11 分钟长片全程是同一个"人"。
- 质检：
- 每段都过"嗡声 / 停顿 / 削波 / 语速"的信号门禁和 ASR 对稿，结尾一个字都不能少；
- 不过关自动换种子；
- 最近一部片子 21 句旁白，经过换种子和一处改写，20 句通过机器对稿，剩下 1 句只差声调，留给人耳确认。
- 速度：M3 Max 上高保真克隆（步数 10）接近实时，生成 1 分钟旁白约 1 分钟。
- 成本：除了电脑和电费，没有按字数、按月的费用；稿子和录音不离开本机。
- 成片：我们用这套流程交付了多部手绘口播视频，比如《用一天，重启你的人生》《错话，该不该翻译？》，最长一部约 11 分钟。
机器质检只能发现明显故障，不能代替听感。每部片子最后一定要完整听一遍。
第十部分：合规——这条比技术更重要
- 只克隆本人或获得授权的声音。 VoxCPM 2 官方明确禁止冒充、诈骗和虚假信息。不要克隆明星、主播、平台音色，也不要用别家 TTS 的输出当参考。
- 发布时标注 AI 生成。 按《人工智能生成合成内容标识办法》，在平台上发布 AI 合成的语音内容，应当主动声明。各平台一般都有"AI 生成"标签，勾上即可。
- 分清三种许可：应用的许可（VoiceStudio 是 AGPL-3.0）、模型的许可（VoxCPM 2 是 Apache-2.0，OmniVoice 是 CC-BY-NC）、声音本身的授权。三者各管各的。
- 我们自己的做法：两款"逼真"声音的原始参考只放在本机个人声音库，不随开源包分发。开源包里附带的是 4 款纯靠文字描述设计出来的合成声音，不涉及任何真人。
常见问题
Q：没有好显卡、也没有 Apple Silicon Mac，能用吗？
能用，但慢：VoiceStudio 0.5.7 有纯 CPU 版本。偶尔做几十秒配音，也可以先用 VoxCPM 官方的在线体验页。注意，在线体验时稿子和录音会上传到远端。
Q：参考音频用 MP3 可以吗？
官方支持 WAV、FLAC、MP3。但要洗参考的话，建议转成 WAV 处理，别只改文件后缀。
Q：我想要四川话、粤语怎么办？
VoxCPM 2 支持 9 种中文方
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
One sentence: VoxCPM 2 is an open-source voice model for OpenBMB, both code and weight are Apache-2.0, commercial. VoiceStudio is an open-source local voiceover desktop application that can install VoxCPM 2 as an engine for it, providing sound libraries, transliteration and local API. Together, they make Chinese bywords close to the real person on their own computers, without paid by word and without uploading articles and recordings.
Elements of this paper
- Syracuse: Why is voiceover so difficult; what is the "preset voice, sound design, voice cloning" for each of the three games; the status of the two projects as of 30 October 2026.
- Again: two routes.
- Route A: VoiceStudio GUI plus VoxCPM 2 engine, zero code.
- Route B: Mac runs VoxCPM 2 directly using mlx-audio, suitable for bulk production and the way we've been doing it.
- The most valuable part: how the reference audio is washed, how the long drafts are organized, how the multilingual and English words are handled, how the quality is tested.
- Curved list: 15 pits per month, each with symptoms, reasons and reconciliation.
- Actual data: Speed and memory on Apple M3 Max, comparison before and after reference audio clean-up, small sample experiment with multi-sounding.
Why do you write this?
Almost all the friends who make videos have suffered on voiceover:
- Make your own recording: a quiet room, a good microphone, a cut. Renumber the entire paragraph.
- Free on-line TTS: When you listen to a machine, you have a flat tone, you have a pause, and the audience cuts out in three seconds.
- Money for members, word for word paid: expensive, nice voice crashes. Your manuscript will be uploaded to the platform and voice will not belong to you.
- Matching: more expensive, longer-term cycles and more waiting for revision.
We almost stepped on all these pits in the open-source project, Sweety-handdraw-video (AI video). The later model was VoiceStudio + VoxCPM 2.
At the hand painting project 2.0 publish, we said, "The highly humanized voiceover is based on this combination," and then a lot of people asked how.
To be precise, the two have been running separately here:
- VoiceStudio: Qwen3-TTS engine, hand painting project 1.0, voiceover and voice transliteration.
- VoxCPM 2 was later added: on Mac, run alone with mlx-udio, became the main parable, creating a unique voiceover.
Now Voice Studio has been able to install VoxCPM 2 as an engine, and I'm going to go over both lines.
Part I: Cope — clarify the concept and cut half the curve.
1. Three methods of AI voiceover
We end up in two combinations: first with "sound design" to find a favorite voice, then with the best reference, and then with "cloning". This does not require real authority, and voice is stable. We'll talk later.
2. VoxCPM 2: One of the most valuable now available open-source Chinese voice models
VoxCPM is an OpenBMB open source voice generation model. VoxCPM 2 on 6 April 2026 publish 2.0. As of this paper, the latest Python package is 2.0.3, publish on May 11, 2026, with --device parameters, and the current decode is faster.
It surprises us most about three patterns of generation and the basis for all the operations that follow:
beginner Note: In the high-pair cloning model, the tone instructions in brackets are ignored. If you want "one voice, another tone", you should use controllable cloning instead of writing instructions in a high-pair mode.
The results on the Seed-TTS-eval are as follows (this is OpenBMB's own publish table, not verified by a third party):
Source: VoxCPM README Performance.
3. VoiceStudio
VoiceStudio is an open source local voiceover desktop application called "Open Source, Local ElevenLabs Alternative" with README in Chinese.
It locates the workstation plus engine manager and does not bind itself to a model:
- Multi-engine: The default engine is OmniVoice, which also contains VoxCPM 2, CosyVoice 3, IndexTTS, MOSS-TTS, GPT-SoVITS, MLX-Audio, etc.
- Sound library: refer to audio recordings, scripts, seeds, next direct selection.
- Auxiliary tools: Voice transliteration (Whisper), human sound separation, video voiceover, audio books and long scripts (support [pause 0.5s] stop sign), sound dictionary.
- Local API: Default Port 3900, with Open AI compatible /v1/udio/peech and /v1/udio/transcriptions, and MCP interfaces with Claude Code, Codex, etc.
The latest version is 0.5.7, 2026. On October 26th, publish, there are several points to note:
- Licence confirmation: Models that do not have an explicit commercial licence, including the default engine OmniVoice, must first receive a licence in the application, otherwise API and MCP calls will fail.
- New installation package and interface: add a pure CPU version, interface redesign.
- paid function: Remote and GPU sharing of these functions is moved to the Pro version of paid.
Why the combination of VoiceStudio + VoxCPM 2?
A fact that many people have not noticed: the VoiceStudio application itself is AGPL-3.0, but its default engine, OmniVoice, has model weights of CC-BY-NC (non-commercial) (model cards).
If you do commercial content such as goods, advertising, knowledge paid, the default engine is not appropriate for licensing. VoxCPM 2 has a weight of Apache-2.0, which is not a concern for commercial purposes, and has strong Chinese and dialect skills.
So the division of labour is:
- VoiceStudio is responsible for: interface, sound library, transliteration, voiceover desk, local API.
- VoxCPM 2 for hearing and commercial.
Reminder: Models allow for commercial use, which does not mean you can clone anyone's voice. Sound authorization is another thing, see the "compliance" section at the end.
Part Two: Our hearts and minds
Before doing business, let's start with how we get to the current programme.
Phase I: tool mix
Each of the earliest projects was divided into two versions: Edge TTS, Piper, Kokoro, Qwen3-TTS, Whisper.cpp, a stand-alone MLX Audio.
- Some Chinese is hard;
- Only a few preset voice;
- Some can be cloned but require separate environments.
Early in October 2026, we resolved to clean up, remove these independent tools and duplicate models, free up about 13 GB disks, and voiceover unified access to VoiceStudio.
Lessons learned: First choose a single portal. The more tools, the more the environment gets, the more time you spend on this.
Stage two: Emotional failure.
After closing, we're going to use the clone model of VoiceStudio plus Qwen3-TTS for the drawing of voiceover: The role of mother and child is to create reference in "sound design" and then clone.
The problem is: Mom's dialogue line needs to be gentle, anxious, happy, but no matter how detailed the instructions, it comes out.
We looked over the source code and found that the Qwen3-TTS cloning branch in VoiceStudio was the realization of mlx-audio, which received emotional instructions but did not transmit them to models at all.
Lessons learned: When emotions don't come out, make sure that the patterns and engines you use do not support emotional instructions. Don't rush prompt. This is one of the reasons why we turned to VoxCPM 2: Its "controllable cloning" clearly supports command tone.
Phase three: VoxCPM 2
We run the 8-bit version of VoxCPM 2 on Mac using mlx-audio. Instead of loading it into VoiceStudio, we built a single Python environment to run. The first cloned male voice, voice, felt better than all previous programmes. We were lazy, and we produced the whole article in one piece, 47 seconds to read it.
It's all in the background. We thought it was the model at first, and we tried to adjust the parameters.
Phase IV: Easing
And we did a spectrum analysis to find the root cause: our seven-second reference tape was cut directly from a background music line.
- Musical residues: 106, 593, 891 Hz on the spectrum.
- Not quiet: the volume at the standstill is only 17.8 dB lower than the volume of the speech, and the normal dry voice should be above 25 dB.
- Slight ripples: the recording has a slight broken sound.
The model treats chorus as "part of this man's voice" and clones it together, and the longer the paragraphs, the bigger.
Reform: Go back to the original recording, pick a full paragraph of 14.9 seconds, ending at the end of the sentence, first with HTDemucs, and then with DeepFilterNet3. As a result, the continuous spectrum disappeared and the pause was lower than the voice 33.5 dB, with a verb error rate of CER 0.035.
On the left side of the chart, there are several lines that run through the entire course, which is the play, and the pause has been "not quiet." On the right, after the clean-up, the line was gone, and the pause was deep down.
Another woman also went through the same clean-up. After the regeneration, our feedback was: "voice approaches, the noise disappears," and the whole thing went through, as the voice of the main girl."
The quality of reference audio determines the ceiling for cloning.
Phase 5: Encyclopedia of long drafts
The next movie is getting longer, from a minute or two to 11 minutes. New questions one by one, followed by article-by-article questions.
- Generation of the whole page: the faster the reading, the occasional buzzing, the inability to stop. It was also officially acknowledged that the long text was unstable and suggested subparagraphs.
- End of sentence: "Try it a few times" and "all" is gone. The reason for that is because we're using the sound of silence at the end of the sentence as the end of the sentence.
- English words don't understand: bugs, Thorsten, the same words, eight seeds, not read at all.
- Machine-testing error: when a long list of men's voices is drawn up, the sonic high will hold steady for several seconds near 120 Hz and be considered a murmur by our humming test. The same sentence: 8 seeds, 7 miscalculated.
We have written the rules into the process on every issue, and the following sections are the rules.
Stage 6: Cloning your voice
We've tried to clone ourselves. From a recording of a meeting I was able to intercept the nine seconds I was talking about, the noise behind the door was off-limits and the error rate was zero, which was technically impeccable.
But it doesn't sound right to put it in a formal intellectual narrative: the sound of everyday speech is not the sound of a voice that is appropriate for byword. Our final division of labour is:
- Population broadcasting, self-messaging: using your own voice.
- Officially speaking: with a more "sounding" man.
Lessons learned: Technical qualifications are not the same as suitability. When you pick the sound, try it with the kind of publish you want, not just "Hello, welcome to my channel."
Phase 7: Multisound
The last time I gave a family badminton video, in 21 words, the ASR tested these wrong:
Our handling:
- Regeneration of seeds;
- It's impossible to rephrase, for example, the opening of the night club;
- There's only a dozen hits left: ASR still listens to actual shots. The difference between the two was that the ASR interpretation itself was unreliable, so it was left to be confirmed that the rest was all cleared.
There's a little experiment that we're doing.
Part III: Operational course A — VoiceStudio + VoxCPM 2 (zero code)
It's for the vast majority of people: I don't want to touch command line, I want an "openable" voiceover software.
Step 0: Make sure your computer's enough.
Disk: VoxCPM 2 engine has its own Python environment. 0.5.6 The installation will first check the availability of approximately 10 GB space and the first generation will also download about 5 GB models. It is suggested to leave at least 20 GB space.
Step 1: Install VoiceStudio
MacOS or Linux, open the terminal, paste:
curl -fsSL https://voicestudio.sh/install | sh
Windows, open PowerShell:
irm https://voicestudio.sh/install.ps1 | iex
You can also go to the Releases page to download the installation package: Mac uses DMG, Windows,.exe. Windows installation packages are not signed, and SmartScreen will play a warning that the source is official GitHub and continue.
The VoiceStudio official provided the prompt "Let Agent help you." Paste the following to Claude Code or Codex:
Install the VoiceStudio Electron app on this device and verify it works, following
https://github.com/debpalash/VoiceStudio/blob/main/docs/install/agent.md
Step 2: Domestic network first mirroring
For the first time, the model is downloaded from Hughes Face, and domestic networks are often slow or unconnected. The VoiceStudio settings have the Hugging Face mirror option, which cuts to hf-mirror.com, which carries its own connectivity test. Get this together, then get the engine. It'll save half the "download" problem.
Step 3: Installation of VoxCPM 2 engine
By official document: Model Catalogue → VoxCPM2 installation (Install). It will be installed in an independent Python environment, without affecting other engines.
- When installed, download the model weights for the first generation, and for the first time slow, patient and so on.
- 0.5.7 Starts with non-commercial models such as the default engine OmniVoice, which require a licence to call. VoxCPM 2 is Apache-2.0, not subject to this restriction.
Mac user note: VoiceStudio also has an "MLX-Audio" engine, which also lists VoxCPM 2. However, using VoiceStudio's own document, reference audio is ignored when called through this engine, and the output is down to 24 kHz. For cloning, please select the VoxCPM2 engine above. Source: VoiceStudio mlx-udio documents.
Step 4: Prepare reference audio (most important, separate sections)
Please jump to the following "Part Five: Reference Audio -- One Hour to Decide Success or Loss" to get the reference tape ready and then come back.
Step 5: Build a Sound Archive
Add a reference recording to the "sound cloning" page, fill out word-by-word versions, that is, every word in the reference tape, word for word and point. Keep a sound file, name it, like "Man-V1"
VoiceStudio rules for processing VoxCPM 2 reference audio, finding out how to reduce the pits:
- It will automatically cut the silence at the end of the reference and then retain up to 30 seconds;
- If it is cut off for more than 30 seconds, it loses the word by word and returns to the ordinary clone because the audio and word by word after the cut is not matched.
So: Reference audio is in 8-15 seconds, do not exceed 30 seconds.
Source: VoiceStudio VoxCPM2 engine document, and 0.5.6 source code.
Step 6: How do three modes trigger in VoiceStudio?
VoiceStudio doesn't let you choose the mode directly, but it's based on what you fill out.
Parameters are recommended as official defaults.
- Guided strength 2.0: lower the natural but possibly run text, higher up the text but easy to sound.
- Steps 10: Officially recommended 4-30, the bigger the detail, the slower. Our main man with 16.
Step 7: Generate, test, read
Upon generation, the audio is transferred back to the text with the VoiceStudio transliteration function:
- Any missing words, multiple words?
- Is the last word at the end?
- Name, number, voice, right?
Change the torrent to regenerate this segment. Note that transliteration only detects obvious problems and ultimately must be heard in the ear.
Step 8: Do not paste a long draft at a time
Either model, long text is easier to read and sound. The correct approach is to create subparagraphs; see Part VI for the rules.
VoiceStudio supports such a stop sign [pause 0.5s] for the "sounding book / long text" function, as well as a dictionary of pronunciation, which can be used uniformly to correct the reading of a word. See Part VII.
Step 9 (step): Generate in local API batch
When VoiceStudio is open, the 3900 port provides the Open AI compatible interface. The parameter name is specified by the interface of 0.5.6:
# List your voice files and get the voice ID
curl -s http://127.0.0.1:3900/v1/audio/voices
# Generate a section of the VoxCPM 2 engine with a specified sound archive White
curl -s http://127.0.0.1:3900/v1/audio/speech \
-H "Content-Type: application/json" \
-d '{
"model": "voxcpm2",
"voice": "Your voice file ID",
"Input": "We have 10 minutes today to make a clear question. " I don't know.
"response_format": "wav",
"seed": 2026,
"num_step": 10,
"guidance_scale": 2.0
}' \
-o part-001.wav
- Seed is important: the same seed, the same parameter, the result can be repeated. If you are not satisfied, change only the seed.
- Format: response format default is mp3, recommended for video.
- Access to Agent: VoiceStudio also provides MCP (http://127.0.0.1:3900/mcp), allowing Claude Code, Codex to call it voiceover directly.
Part IV: Operational course B - Mac Run directly on VoxCPM 2
They are suitable for people who want to produce in bulk and who want to connect voiceover to their own process. This is how our hand-painted video project works: automatic generation of each sidebar, automatic mass checking, non-cleared automatic re-entry of seeds, which is then spelled into a complete verb belt.
1. Installation
python3 -m venv ~/voxcpm2-mlx
source ~/voxcpm2-mlx/bin/activate
pip install -U mlx-audio soundfile mlx-whisper
# The domestic network can be mirrored
export HF_ENDPOINT=https://hf-mirror.com
mlx-audio (MIT permission) started supporting VoxCPM 2 from 0.4.4 and as of this time is 0.6.3. Our production environment is 0.5.8 and the following code has been validated on 0.5.8; any change in the name of the new version of the parameter is subject to the official description.
There are three sets of weights:
2. Next to Kopo-Klon White
import mlx.core as mx
import numpy as np
import soundfile as sf
from mlx_audio.tts.utils import load
model = load("mlx-community/VoxCPM2-8bit")
== sync, corrected by elderman ==
# Word by word Draft
text = "We'll take 10 minutes today to get this whole question straight."
mx.random.seed (2026) # Fixed seed, results repeated
chunks = list(model.generate(
text=text,
Ref udio=ref, prompt audio=ref, prompt text=ref text, # All three = high-precision cloning
inference_timesteps=10, cfg_value=2.0,
max tokens=len(text)* 6 + 10, # upper limit of official "text length x 6 + 10" to prevent discontinuation
))
audio = np.concatenate([np.asarray(c.audio) for c in chunks])
sf.write("part-001.wav", audio, chunks[0].sample_rate) # 48 kHz
Two other models:
- Controllable cloning: only ref audio, plus instruct= "Slower speech, speech temperature and."
- Sound design: Not for reference, only for instract = "Middle-aged man, low-magnetic, steady, knowledgeable."
3. "Record" immediately after generation
import mlx_whisper
r = mlx_whisper.transcribe("part-001.wav",
path_or_hf_repo="mlx-community/whisper-large-v3-turbo",
language="zh")
print(r["text"])
Compare the result of the identification to the original text word by word, calculate the error rate:
- More than 0.15, or the end is missing: the paragraph is invalidated and the seeds are replaced.
- Auto-Z97Z up to three times.
4. Measurement velocity and memory
Conditions: Apple M3 Max, 36 GB RAM, VoxCPM2-8bit, step 10.
The first load model is about 19 seconds.
convert: Our main voice is about 6.6 words/s, a 4581 word piece, about 708 seconds (11 minutes 48 seconds).
Don't start two processes in parallel: we tried two generation processes at the same time, but it took longer, and one time we were systematically killed because our memory was tight. Single processes are the most stable in order.
Part V: Reference audio
This is the most valuable section of the whole. The quality of cloning 80 per cent is determined by reference.
1. Standards for reference audio
2. Where does the sound come from? Three sources of compliance
- Make your own recording: Put your cell phone by the quiet closet, 20 centimeters off your mouth, read a 15-second speech in a normal speech.
- Recording of authorized persons: a clear authorization, preferably in writing.
- Create one with the sound design: this is our most recommended approach, which involves no real person, but can also stabilize voice:
- Writing sound descriptions, the official suggestion is "identity + texture + scene", for example: "Middle-aged man, low, magnetic, strong, robust, knowledge- and story-telling."
- The same description produces four times, different seeds.
- Pick one of the cleanest, most like you want, for reference.
- voice fixed it when it was all cloned.
The sound description only features the sound, does not write "like some anchor" or "some voice" on a platform, and does not use the output of other TVS as a reference.
What's the reference to dirty? Wash
If you have a tape of music or noise:
Graphical interface: VoiceStudio has a voice separation and audio cleanup function, which goes through the recording first.
command line: Our approach is two steps, first to separate the voices and then to denominate the noise.
#1) Intercept 8-15 seconds and end of sentence (unio, 48 kHz)
ffmpeg -ss 31.6 -to 46.0 -i source.wav -ac 1 -ar 48000 cut.wav
#2) Separating the human voice (Demucs, MIT permission), retaining only the human voice
pip install demucs
# result in separated/htdemucs/cut/vocals.wav
#3) Noise DeepFilterNet
pip install deepfilternet
deepFilter separated/htdemucs/cut/vocals.wav -o cleaned/
How do you know if it's clean or not?
- It's the simplest thing: Open it with software like Audacity, cut to spectrum. If you see a line running through the entire course, there's a mix or a current sound.
- Put on the headset stop: the pause should be dead, not something.
- Our quantization gate is closed:
- The pause is 25 dB or more and a clean reference is below -30 dB;
- No continuous single frequency lines;
- Can't chop.
VoxCPM 2 itself has denoise parameters, but it only handles reference audio and follows 16 kHz pipe lines, which officially say "slightly changes voice". Keep it closed if it's clean.
Part VI: Production of long drafts — segment, standstill, end of sentence
1. Sub-rule
2. Length of stay between paragraphs Hey.
Inserts a pure silence when you spell. Here's a set of what we've learned from acceptance check's film, as well as a long story:
3. At the end of the sentence
Once the model is finished, there's a little more noise, like a breath, half a noise. Our handling:
- Look back at the end of the last word, the first part lasts 30 milliseconds below -40 dB, the true end of the sentence.
- Don't start with words: b, t, g, before a sonar mother opens, his lips or tongue will be closed for dozens of milliseconds, and that silence is not the end of the sentence. That's why we lost the word "a few more times."
- Keep 0.12 seconds after the end, and use at least 50 milliseconds to fade out.
- Once again: confirm that the last word is still there.
4. Mixing
- The overall volume of the narration is around -15 LuFS (the range commonly used on the short video platform).
- Background music is down when talking.
- Total peak not exceeding -1.5 dBTP.
Part VII: Multisound, English and Numerical - 3 hard bones of voiceover in Chinese
Multisound: We did a little experiment.
The official method of giving is to write the accents in the text, such as {ni3}{hao3} and turn off the normalize.
There is also an earth-based approach in the community: homonymization. subtitle is still in the original language. VoiceStudio's Dictionary supports word substitution, which automates the matter.
We tried five seeds in our own environment. The environment is mlx-audio 0.5.8, 8-bit, high-punct cloning, resulting in a whisper interpretation:
A few notes:
- sample is small, just 5 seeds, conclusions for reference only.
- MLX not necessarily fully supports official spelling. An official PyTorch version of a friend can give priority to the spelling label.
- ASR sound itself unreliable. Whisper gave two results in both ways, and we couldn't judge it, so we didn't put it in the table. Finally, we have to be confirmed.
Our practice now:
- Avoid writing: the words that are easy to read, the words change.
- Can't replace it, replace it with the same words and retain the original text in subtitle.
- Paragraph-by-paragraph ASR text-reading, changing seeds by error, up to three times.
- Not yet. Change that sentence.
- Last person heard it.
2. English language: minimum use
The Chinese voice model reads Chinese in English and often reads "Chinese recognition, English ears don't understand." We measured the main man (VoxCPM2-8bit cloning):
- bug, Thorsten: Eight seeds in the same sentence, not once.
-Guit, Amp: Some seeds can hear.
- AI, Agent, Token: Most can hear.
So we set the rules:
- In Chinese, the common word, like bug, is "problem".
- A common Chinese translation of a person ' s name may be retained in English.
- The English language must be preserved, produced and individually cut out, identified and heard again in English and replaced with seeds without hearing, up to six times.
3. Numbers, dates, units: In Chinese
The official normalize parameter can handle digital reading, but the MLX ported version does not. The surest way to do it is to write it directly in your speech:
- "2026" to "2026"
- "1024" to read "104" or "1024"?
- "2.0.3" to "2:03".
Part VIII: 15 pits we stepped on
There was also an uncharted one: the cloning of one ' s voice was technically full, not necessarily for official content. The sound is selected with real content.
Part IX: Final effects
- voice: Two major voices (a man and a woman) were cleared and the noise disappeared. "voice approaches, the murmur disappears," and the whole thing passes.
- Stability: the same reference plus fixed seeds, the results can be repeated. Eleven minutes long film is the same person.
- Examination:
- Each passage has a signal door-bar and ASR text-reading that ends with one word;
- No automatic exchange of seeds for cleared;
- A recent film, 21 words, changed seeds and a rewrite, 20 words through machines, and one sentence left behind, with an ear for confirmation.
- Speed: The M3 Max high-punct cloning (steps 10) is close to real time, generating about 1 minute by 1 minute.
- Cost: No word, monthly costs except for computers and electricity; articles and recordings do not leave the machine.
- Make a film: We used this process to deliver multiple hand-drawing videos, such as "A Day Restart Your Life, Should I Translator?" for about 11 minutes.
Machine quality checks can only detect obvious malfunctions and cannot replace hearing. Each film must be heard in its entirety.
Part X: Compliance - this is more important than technology
- To clone the voice of the person or of the authorized person. VoxCPM 2 The official prohibition of impersonation, fraud and false information is clear. Do not clone stars, anchors, platforms voice, or use other TVS outputs as reference.
- publish timemark AI generation. Based on the Artificial Intelligence Synthetic Content Identification Scheme, speech content synthesized from publish AI on the platform should be declared on its own initiative. Each platform usually has an "AI Generating" label.
- Distinguishing three types of licences: the licence for the application (VoiceStudio is AGPL-3.0), the licence for the model (VoxCPM 2 is Apache-2.0, OmniVoice is CC-BY-NC), and the authorization for the sound itself. All three are in charge.
- Our own approach: the original references to two "real" sounds are only in the personal voice library of the machine and are not distributed with the source pack. The open source package is accompanied by 4 sections of a synthetic sound designed purely by a description of the text, not involving any real person.
Common problems
Q: No good GPU or Apple Silicon Mac, can it work?
Available, but slow: VoiceStudio 0.5.7 has a pure CPU version. Do a couple of seconds of voiceover, or start with the VoxCPM official online experience page. Note that online experience time articles and audio recording sessions reach the remote end.
Q: Reference Audio MP3 OK?
Official support for WAV, FLAC, MP3. However, for reference purposes, it is suggested that it be converted to WAV processing, rather than simply suffixing files.
Q: What about Sichuan and Chinese?
VoxCPM 2 supports 9 Chinese
Auto-fetched and translated by AI Atlas. All rights belong to the original author.