TTS

Signal Desk

modal-lab · 025–034 Listening Room
Live · Modal smoke · 2026-08-11

把 TTS 用量榜 S+A 放在一张工作台上听完

Kokoro → Higgs（025–034）。30 条 Modal 实测音频全部上架：波形可视化、成本 / VRAM 条形图、实时频谱，以及可筛选、可搜索的试听列。空格播放 / 暂停。

开始试听 30 条 打开实测矩阵 返回 lab 总览
30smoke 音频 · 全部展示
10S+A 模型 · 025–034
$0.001最便宜 · Kokoro T4
22 GB峰值 VRAM · Fish

模型入口

025 · S1

Kokoro-82M

HF 下载榜 #1 · 最便宜基线

GPU T4 VRAM 0.8GB ~$0.001 2 tracks
打开专页 → 026 · S2

Chatterbox

多语 + Turbo 克隆 · MIT

GPU L4 VRAM 3.5GB ~$0.014 3 tracks
打开专页 → 027 · S3

Qwen3-TTS

Custom / Design / Clone

GPU L4 VRAM 4.5GB ~$0.005 4 tracks
打开专页 → 028 · S4

Fish S2 Pro

AA Elo 开源 #1 · Research

GPU L40S VRAM 22GB ~$0.06 4 tracks
打开专页 → 029 · A1

VoxCPM2

Tokenizer-free 2B · 速度/延迟

GPU L4 VRAM 5.8GB ~$0.007 4 tracks
打开专页 → 030 · A2

VibeVoice RT

GH stars #1 · 流式长文

GPU L4 VRAM 2.8GB ~$0.005 3 tracks
打开专页 → 031 · A3

CosyVoice3

中文方言 SOTA · Apache

GPU L4 VRAM 3.6GB ~$0.007 3 tracks
打开专页 → 032 · A4

IndexTTS-2

时长 + 情感 · 配音刚需

GPU L4 VRAM 7.3GB ~$0.008 3 tracks
打开专页 → 033 · A5

F5-TTS

零样本最便宜克隆档

GPU L4 VRAM 2.1GB ~$0.0025 2 tracks
打开专页 → 034 · A6

Higgs v2

场景/表现力 · TTS 收官

GPU L40S VRAM 15.7GB ~$0.011 2 tracks
打开专页 →

成本对照 · 模型均费

各模型 smoke est USD 均值（含冷启动墙钟）

Kokoro-82M
$0.0014
F5-TTS
$0.0026
Qwen3-TTS
$0.0049
VibeVoice R
$0.0055
CosyVoice3
$0.0071
IndexTTS-2
$0.0082
VoxCPM2
$0.0093
Chatterbox
$0.0144
Higgs v2
$0.0192
Fish S2 Pro
$0.062

峰值 VRAM

决定最低可用卡型 · Fish / Higgs 需 L40S 档

Kokoro-82M
0.8G
F5-TTS
2.1G
VibeVoice R
2.8G
Chatterbox
3.6G
CosyVoice3
3.6G
Qwen3-TTS
4.7G
VoxCPM2
5.8G
IndexTTS-2
7.4G
Higgs v2
15.7G
Fish S2 Pro
22.2G

全部 30 条结果

点播放 · 波形来自真实 wav · 底部 dock 带频谱 · Space ← →

30 / 30
没有匹配的结果，试试清空筛选或搜索词。

025 · Kokoro-82M

S1 · 2 tracks
Kokoro-82M

EN · af_heart

v1 · 24 kHz · Apache-2.0

Kokoro is an open-weight text-to-speech model with eighty-two million parameters. Despite its lightweight architecture, it delivers strong quality while staying fast and cheap.
GPU T4 VRAM 0.8G wall 7.4s gen 2.0s dur 11.1s usd $0.0012
Kokoro-82M

ZH · zf_001

v1.1-zh · 24 kHz · Apache-2.0

你好，这是 modal-lab 第零二五号实验。Kokoro 是一款参数量仅八千两百万的开源语音合成模型，体积小、速度快，适合作为成本基线。
GPU T4 VRAM 0.8G wall 9.9s gen 2.7s dur 10.9s usd $0.0016

026 · Chatterbox

S2 · 3 tracks
Chatterbox

MTL EN

Multilingual V3

Chatterbox is running on Modal. This is the multilingual model — natural speech across many languages.
GPU L4 VRAM 3.4G wall 64.5s gen 21.9s dur 5.6s usd $0.014
Chatterbox

MTL ZH

Multilingual V3

你好，今天天气真不错，希望你有一个愉快的周末。这是 modal-lab 第零二六号 Chatterbox 实验。
GPU L4 VRAM 3.6G wall 69.0s gen 28.6s dur 12.1s usd $0.015
Chatterbox

Turbo · Lucy clone

Turbo + [chuckle]

Hi there, Sarah here from MochaFone calling you back [chuckle], have you got one minute to chat about the billing issue?
GPU L4 VRAM 3.3G wall 61.7s gen 27.5s dur 7.2s usd $0.014

027 · Qwen3-TTS

S3 · 4 tracks
Qwen3-TTS

Custom · Vivian ZH

CustomVoice 1.7B

你好，这是 modal-lab 第零二七号 Qwen3-TTS 实验。 · instruct: 用自然、温和的语气说
GPU L4 VRAM 4.5G wall 24.0s gen 16.8s dur 9.3s usd $0.0053
Qwen3-TTS

Custom · Ryan EN

CustomVoice 1.7B

Hello from Modal lab experiment zero two seven. · instruct: Speak cheerfully and clearly.
GPU L4 VRAM 4.4G wall 18.8s gen 13.2s dur 7.7s usd $0.0042
Qwen3-TTS

VoiceDesign ZH

文本造声

哥哥，你回来啦，人家等了你好久好久了，要抱抱！ · design: 撒娇稚嫩萝莉女声
GPU L4 VRAM 4.3G wall 18.4s gen 11.5s dur 5.4s usd $0.0041
Qwen3-TTS

Clone EN

3s 克隆

I am solving the equation on the board.
GPU L4 VRAM 4.7G wall 26.4s gen 19.7s dur 6.6s usd $0.0059

028 · Fish S2 Pro

S4 · 4 tracks
Fish S2 Pro

EN · random voice

S2 Pro random

Fish Audio S2 Pro is running on Modal. This is the open-weights quality leader — natural multilingual speech.
GPU L40S VRAM 22.2G wall 131.1s gen 58.0s dur 8.9s usd $0.071
Fish S2 Pro

ZH · random voice

S2 Pro random

你好，这是 modal-lab 第零二八号 Fish Audio S2 Pro 实验。今天天气真不错，希望你有一个愉快的周末。
GPU L40S VRAM 22.2G wall 114.7s gen 59.3s dur 8.6s usd $0.062
Fish S2 Pro

Inline tags

[emotion] tags

Hi there [excited], welcome to the demo! [chuckle] Can you believe this voice has over fifteen thousand control tags? [whisper] Soft parts work too.
GPU L40S VRAM 22.2G wall 123.7s gen 69.3s dur 10.1s usd $0.067
Fish S2 Pro

Clone EN

reference clone

I am solving the equation on the board. Nobody said it would be this hard, but we are almost there!
GPU L40S VRAM 22.2G wall 88.0s gen 32.7s dur 5.6s usd $0.048

029 · VoxCPM2

A1 · 4 tracks
VoxCPM2

Smoke EN

tokenizer-free TTS

VoxCPM2 is running on Modal. This is a tokenizer-free multilingual text to speech model optimized for speed and natural prosody.
GPU L4 VRAM 5.8G wall 33.0s gen 13.1s dur 8.5s usd $0.0073
VoxCPM2

Smoke ZH

tokenizer-free TTS

你好，这是 modal-lab 第零二九号 VoxCPM2 实验。今天天气真不错，希望你有一个愉快的周末。
GPU L4 VRAM 5.8G wall 32.0s gen 13.9s dur 9.3s usd $0.0071
VoxCPM2

Voice Design

文本造声

(A young woman, gentle and sweet voice, warm midrange)Hello, welcome to VoxCPM2 voice design on Modal lab.
GPU L4 VRAM 5.6G wall 34.8s gen 9.6s dur 4.0s usd $0.0077
VoxCPM2

Clone · reference_speaker

48k 可控克隆

This is a controllable voice clone generated by VoxCPM2 on Modal.
GPU L4 VRAM 5.8G wall 67.0s gen 39.7s dur 4.2s usd $0.015

030 · VibeVoice RT

A2 · 3 tracks
VibeVoice RT

Carter EN

Realtime · man preset

VibeVoice Realtime is running on Modal. This lightweight model supports streaming text input and robust long-form speech generation.
GPU L4 VRAM 2.8G wall 25.4s gen 8.7s dur 8.1s usd $0.0056
VibeVoice RT

Emma Long

~27s 长句 · RTF≈0.6

Welcome to modal-lab experiment zero three zero. VibeVoice was designed for long conversational audio and podcast narration. Even this realtime half billion parameter variant can keep going for minutes while remaining lightweight enough for a single L4 GPU. Today we measure cold-
GPU L4 VRAM 2.8G wall 26.1s gen 16.3s dur 27.2s usd $0.0058
VibeVoice RT

Emma Short

Realtime · woman preset

Hello, this is Emma speaking through VibeVoice Realtime on Modal. Switching speakers only requires a different voice preset cache.
GPU L4 VRAM 2.8G wall 22.4s gen 8.3s dur 8.7s usd $0.0050

031 · CosyVoice3

A3 · 3 tracks
CosyVoice3

Zero-shot ZH

Fun-CosyVoice3

你好，这是 modal-lab 第零三一号 CosyVoice3 实验。今天天气真不错，希望你有一个愉快的周末。
GPU L4 VRAM 3.6G wall 42.2s gen 13.2s dur 9.4s usd $0.0094
CosyVoice3

四川话 instruct

instruct2 方言

收到好友从远方寄来的生日礼物，那份意外的惊喜与深深的祝福让我心中充满了甜蜜的快乐。
GPU L4 VRAM 3.6G wall 21.9s gen 8.8s dur 8.8s usd $0.0049
CosyVoice3

Cross-lingual EN

跨语种

Hello from CosyVoice3 on Modal lab. This is a zero-shot multilingual speech synthesis demo.
GPU L4 VRAM 3.6G wall 31.2s gen 10.0s dur 7.2s usd $0.0069

032 · IndexTTS-2

A4 · 3 tracks
IndexTTS-2

Zero-shot ZH

IndexTTS-2

你好，这是 modal-lab 第零三二号 IndexTTS-2 实验。精确时长与情感控制，是中文配音的硬通货。
GPU L4 VRAM 7.4G wall 33.0s gen 13.1s dur 8.3s usd $0.0073
IndexTTS-2

Zero-shot EN

IndexTTS-2

Hello from IndexTTS-2 on Modal lab. Industrial zero-shot speech with duration and emotion control.
GPU L4 VRAM 7.3G wall 36.9s gen 14.6s dur 7.6s usd $0.0082
IndexTTS-2

Emo · 悲伤

emo_text

这些年的时光终究是错付了……我只希望你能过得比我好一点。
GPU L4 VRAM 7.3G wall 41.4s gen 16.0s dur 8.3s usd $0.0092

033 · F5-TTS

A5 · 2 tracks
F5-TTS

Clone EN

F5TTS_v1_Base

Hello from F5-TTS on Modal lab. This is a zero-shot voice clone demo with flow matching.
GPU L4 VRAM 2.1G wall 12.0s gen 5.7s dur 8.4s usd $0.0027
F5-TTS

Clone ZH

F5TTS_v1_Base

你好，这是 modal-lab 第零三三号 F5-TTS 实验。零样本克隆，中英皆可。
GPU L4 VRAM 2.1G wall 11.2s gen 6.2s dur 11.0s usd $0.0025

034 · Higgs v2

A6 · 2 tracks
Higgs v2

Quiet room EN

scene prompt

The sun rises in the east and sets in the west. This simple fact has been observed by humans for thousands of years.
GPU L40S VRAM 15.7G wall 49.6s gen 6.2s dur 8.3s usd $0.027
Higgs v2

Expressive

excited scene

Wow! Can you believe it? We finally shipped the last TTS experiment — Higgs Audio on Modal lab. What a journey!
GPU L40S VRAM 15.7G wall 21.1s gen 7.0s dur 9.9s usd $0.011

实测矩阵

模型样本GPUVRAMwall sgen saudio sest USD
Kokoro-82M EN · af_heart T4 0.8 7.4 2.0 11.1 $0.0012
Kokoro-82M ZH · zf_001 T4 0.8 9.9 2.7 10.9 $0.0016
Chatterbox MTL EN L4 3.4 64.5 21.9 5.6 $0.0143
Chatterbox MTL ZH L4 3.6 69.0 28.6 12.1 $0.0153
Chatterbox Turbo · Lucy clone L4 3.3 61.7 27.5 7.2 $0.0137
Qwen3-TTS Custom · Vivian ZH L4 4.5 24.0 16.8 9.3 $0.0053
Qwen3-TTS Custom · Ryan EN L4 4.4 18.8 13.2 7.7 $0.0042
Qwen3-TTS VoiceDesign ZH L4 4.3 18.4 11.5 5.4 $0.0041
Qwen3-TTS Clone EN L4 4.7 26.4 19.7 6.6 $0.0059
Fish S2 Pro EN · random voice L40S 22.2 131.1 58.0 8.9 $0.071
Fish S2 Pro ZH · random voice L40S 22.2 114.7 59.3 8.6 $0.0622
Fish S2 Pro Inline tags L40S 22.2 123.7 69.3 10.1 $0.067
Fish S2 Pro Clone EN L40S 22.2 88.0 32.7 5.6 $0.0477
VoxCPM2 Smoke EN L4 5.8 33.0 13.1 8.5 $0.0073
VoxCPM2 Smoke ZH L4 5.8 32.0 13.9 9.3 $0.0071
VoxCPM2 Voice Design L4 5.6 34.8 9.6 4.0 $0.0077
VoxCPM2 Clone · reference_speaker L4 5.8 67.0 39.7 4.2 $0.0149
VibeVoice RT Carter EN L4 2.8 25.4 8.7 8.1 $0.0056
VibeVoice RT Emma Long L4 2.8 26.1 16.3 27.2 $0.0058
VibeVoice RT Emma Short L4 2.8 22.4 8.3 8.7 $0.0050
CosyVoice3 Zero-shot ZH L4 3.6 42.2 13.2 9.4 $0.0094
CosyVoice3 四川话 instruct L4 3.6 21.9 8.8 8.8 $0.0049
CosyVoice3 Cross-lingual EN L4 3.6 31.2 10.0 7.2 $0.0069
IndexTTS-2 Zero-shot ZH L4 7.4 33.0 13.1 8.3 $0.0073
IndexTTS-2 Zero-shot EN L4 7.3 36.9 14.6 7.6 $0.0082
IndexTTS-2 Emo · 悲伤 L4 7.3 41.4 16.0 8.3 $0.0092
F5-TTS Clone EN L4 2.1 12.0 5.7 8.4 $0.0027
F5-TTS Clone ZH L4 2.1 11.2 6.2 11.0 $0.0025
Higgs v2 Quiet room EN L40S 15.7 49.6 6.2 8.3 $0.0269
Higgs v2 Expressive L40S 15.7 21.1 7.0 9.9 $0.0114