Hugging Face's logo

Spaces:
MiniMaxAI
/
MiniMax-Speech-Tech-Report
Running

App Files Community
2
MiniMax-Speech-Tech-Report / index.html
jiaxin
feat: add citation
67e8274
raw
history blame contribute delete
41.7 kB
 <!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>MiniMax-Speech Tech Report | Intrinsic Zero-Shot Text-to-Speech with a Learnable Speaker Encoder</title>
<meta name="description"
content=" MiniMax-Speech, an autoregressive Transformer-based Text-to-Speech (TTS) model that generates high-quality speech" />
<meta name="keywords" content="latex.css,css library,class-less css,latex css" />
<meta property="og:title"
content="MiniMax-Speech Tech Report | Intrinsic Zero-Shot Text-to-Speech with a Learnable Speaker Encoder" />
<meta property="og:url" content="https://minimax-ai.github.io/tts_tech_report" />
<meta property="og:description"
content=" MiniMax-Speech, an autoregressive Transformer-based Text-to-Speech (TTS) model that generates high-quality speech" />
<meta property="og:type" content="website" />
<link rel="stylesheet" href="style.css" />
</head>
<body id="top" class="text-justify">
<header
style="background-image: url('assets/images/header-bg.jpeg'); background-size: cover; background-position: center; padding: 1rem 0; border-radius: 1rem;">
<h1>MiniMax-Speech</h1>
<h4 style="font-size: 1.3rem; line-height: 1; text-align: center;">Intrinsic Zero-Shot Text-to-Speech
with a
Learnable Speaker
Encoder</h4>
<p class="author">
MiniMax Team <span class="date">May 2025</span><br />
<a style="font-size: 1.1rem;" target="_blank" href="https://arxiv.org/abs/2505.07916">[Tech
Report]</a>
<a style="font-size: 1.1rem; margin-left: 1rem;" target="_blank"
href="https://huggingface.co/datasets/MiniMaxAI/TTS-Multilingual-Test-Set">[Multilingual Test Set]</a>
<a style="font-size: 1.1rem; margin-left: 1rem;" target="_blank" href="https://github.com/MiniMax-AI">[GitHub]</a>
</p>
</header>
<div class="abstract">
<h2>Abstract</h2>
<p style="text-align: left;">
We introduce MiniMax-Speech, an autoregressive Transformer-based Text-to-Speech (TTS) model that generates
high-quality
speech. A key innovation is our learnable speaker encoder, which extracts timbre features from a reference audio
without
requiring its transcription. This enables MiniMax-Speech to produce highly expressive speech with timbre
consistent with
the reference in a zero-shot manner, while also supporting one-shot voice cloning with exceptionally high
similarity to
the reference voice. In addition, the overall quality of the synthesized audio is enhanced through the proposed
Flow-VAE. Our model supports 32 languages and demonstrates excellent performance across multiple objective and
subjective evaluations metrics. Notably, it achieves state-of-the-art (SOTA) results on objective voice cloning
metrics
(Word Error Rate and Speaker Similarity) and has secured the top position on the public TTS Arena leaderboard.
Another
key strength of MiniMax-Speech, granted by the robust and disentangled representations from the speaker encoder,
is its
extensibility without modifying the base model, enabling various applications such as: arbitrary voice emotion
control
via LoRA; text to voice (T2V) by synthesizing timbre features directly from text description; and professional
voice
cloning (PVC) by fine-tuning timbre features with additional data.
</p>
</div>
<nav role="navigation" class="toc">
<h2>Explore MiniMax-Speech</h2>
<p>Welcome to visit
<a href="https://www.minimax.io/audio">MiniMax Audio</a> and
explore our powerful TTS features.
</p>
<h2>Contents</h2>
<ol>
<li>
<a href="#architecture-overview">Architecture Overview</a>
</li>
<li>
<a href="#expressiveness-demonstrations">Expressiveness Demonstrations</a>
<ol>
<li><a href="#showcase-with-high-versatility">Showcase with High Versatility</a></li>
<li><a href="#showcase-with-multiple-generation-attempts">Showcase with Multiple Generation Attempts</a></li>
</ol>
</li>
<li><a href="#zero-shot-vs-one-shot-demonstrations">Zero-Shot vs. One-Shot Demonstrations</a></li>
<li><a href="#multilingual-and-cross-lingual-capabilities-demonstrations">Multilingual and Cross-Lingual
Capabilities Demonstrations</a></li>
<li><a href="#flow-vae-vs-vae-comparisons">Flow-VAE vs. VAE Comparisons</a></li>
<li><a href="#professional-voice-clone-pvc-demonstrations">Professional Voice Clone (PVC) Demonstrations</a></li>
<li><a href="#emotion-control-demonstrations">Emotion Control Demonstrations</a></li>
<li><a href="#text-prompted-voice-generation-demonstrations">Text-Prompted Voice Generation Demonstrations</a>
</li>
<li><a href="#comparison-of-voice-naturalness">Comparison of voice
naturalness with the previous generation products</a></li>
<li><a href="#citation">Citation</a></li>
</ol>
</nav>
<main>
<article>
<div class="article-block">
<h2 id="architecture-overview">Architecture Overview</h2>
<figure>
<img src="assets/images/system-overview.jpg" loading="lazy" alt="System Architecture" width="100%"
height="auto" />
<figcaption>
An overview of the architecture of MiniMax-Speech.
</figcaption>
</figure>
</div>
<div class="article-block">
<h2 id="expressiveness-demonstrations">Expressiveness Demonstrations</h2>
<h3 id="showcase-with-high-versatility">Showcase with High Versatility</h3>
<div class="scroll-wrapper">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col" style="width: 40%;">Description</th>
<th scope="col" style="width: 30%; text-align: center;">Source Audio</th>
<th scope="col" style="width: 30%; text-align: center;">Generated Audio</th>
</tr>
<tr class="border-bottom-thin">
<td>
A Compelling and Persuasive Speaker Voice
</td>
<td>
<audio class="audio-md" src="assets/audios/Marketing_Voice_Sourse.wav" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Compelling%20and%20Persuasive.wav" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
A Clear and Explanatory Voice with Broad Emotional Dynamics Across Different Texts
</td>
<td>
<audio class="audio-md" src="assets/audios/Science_Voice_Sourse.wav" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Explanatory%20Broad%20Emotional.wav" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
Another Explanatory Voice with Supernatural Prosody, <br>
Featuring Distinct Ethnic and Age Characteristics
</td>
<td>
<audio class="audio-md" src="assets/audios/Sociology_Sourse.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Explanatory Supernatural Prosody.MP3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
A Warm and Magnetic Voice that Brings Comfort
</td>
<td>
<audio class="audio-md" src="assets/audios/Warm%20and%20Magnetic_Sourse.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Warm%20and%20Magnetic.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
An ASMR Whispering Voice with Generated Breathing and Sound Effects
</td>
<td>
<audio class="audio-md" src="assets/audios/Breathy%20ASMR_Sourse.wav" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Breathy%20ASMR.MP3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
A Robotic Voice with Rich Bass Resonance and Spatial Presence
</td>
<td>
<audio class="audio-md" src="assets/audios/Lucky%20Robot_Sourse.wav" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Lucky%20Robot.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
A Sardonic Mature Female Voice
</td>
<td>
<audio class="audio-md" src="assets/audios/Onee-san_Sourse.MP3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Onee-san.wav" controls></audio>
</td>
</tr>
</tbody>
</table>
</div>
<h3 id="showcase-with-multiple-generation-attempts">Showcase with Multiple Generation Attempts, Post-Processing
Audio Effects and Added Sound Effects</h3>
<div class="scroll-wrapper">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col" style="width: 50%;">Description</th>
<th scope="col" style="width: 50%; text-align: center;">Generated Audio</th>
</tr>
<tr class="border-bottom-thin">
<td>
A Husky Male Voice: From Soft Murmur to Excitement to Anger, then to Whispers
</td>
<td>
<audio class="audio-lg" src="assets/audios/Murmur-Excitement-Anger-%20Whispers.MP3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
An Angry Female Voice: From Soft Murmur to Rage to Reminiscence, then to Weeping
</td>
<td>
<audio class="audio-lg" src="assets/audios/Neutral-Rage-Reminiscence-Weeping.MP3" controls></audio>
</td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="article-block">
<h2 id="zero-shot-vs-one-shot-demonstrations">Zero-Shot vs. One-Shot Demonstrations</h2>
<p>
ZeroShot maintains speaker identity while generating more natural emotions, pauses, and other expressive
features based
on the text content, whereas OneShot adheres more strictly to the speaker characteristics (prosody, speech
rate,
emotions, etc.). For details of Zero-Shot and One-Shot, refer to the <a
href="https://arxiv.org/abs/2505.07916" target="_blank">technical report</a>.
</p>
<div class="scroll-wrapper" style="margin-top: 2rem;">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col">Source Audio</th>
<th scope="col">Text</th>
<th scope="col">Zero-Shot Version</th>
<th scope="col">One-Shot Version</th>
<th scope="col">Elevenlabs Multilingual_v2</th>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-sm" src="assets/audios/Lyrical%20Cantonese_Prompt.WAV" controls></audio>
</td>
<td>
命运就算颠沛流离，<br>
命运就算曲折离奇，<br>
命运就算恐吓着你，<br>
做人没趣味。<br>
别流泪，心酸，更不应舍弃。<br>
我愿能，一生永远陪伴你。
</td>
<td>
<audio class="audio-sm" src="assets/audios/Lyrical%20Cantonese_ZeroShot.mp3" controls></audio>
Preserving Distinctive Voice<br>
Timbre and Expressive <br>
Prosody with Regularized <br>
Pausing and Speech Rate
</td>
<td>
<audio class="audio-sm" src="assets/audios/Lyrical%20Cantonese_Oneshot.mp3" controls></audio>
Better Reproduction of<br>
Prompt's Exaggerated Speech<br>
Rate and Characteristic<br>
Phrase-Initial Pauses
</td>
<td>
Cantonese not supported
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-sm" src="assets/audios/Breaking%20Down%20Mandarin_Prompt.WAV" controls></audio>
</td>
<td>
你们这些躲在道德高地的懦夫，<br>
敢承认自己对本我的恐惧吗？<br>
回答我！嗯？你回答我！<br>
Look in my eyes！<br>
老子写梦的解析时<br>
你们还在玩泥巴，<br>
我精神分析引论每个字母都能<br>
刺穿文明社会的虚伪面具，<br>
我解剖潜意识就像<br>
外科医生划开皮肤。<br>
是不是啊？说话！
</td>
<td>
<audio class="audio-sm" src="assets/audios/Breaking%20Down%20Mandarin_ZeroShot.mp3" controls></audio>
Capable of Generating<br>
Relatively Calmer Emotions<br>
while Preserving Voice<br>
Identity
</td>
<td>
<audio class="audio-sm" src="assets/audios/Breaking%20Down%20Mandarin_OneShot.mp3" controls></audio>
Consistently Reproducing the<br>
Angry Emotion from Prompt<br>
in Every Utterance
</td>
<td>
<audio class="audio-sm" src="assets/audios/ElevenLabs_Breaking Down Mandarin.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-sm" src="assets/audios/Quirky%20Female%20English_Prompt.MP3" controls></audio>
</td>
<td>
Would you believe what happened at the<br>
grocery store today? My goodness! The<br>
avocados were on sale - half price! Half<br>
price! I bought twenty of them!
</td>
<td>
<audio class="audio-sm" src="assets/audios/Quirky%20Female%20English_ZeroShot.MP3" controls></audio>
Effectively follows textual cues<br>
for both longer and shorter<br>
inter-sentence pauses
</td>
<td>
<audio class="audio-sm" src="assets/audios/Quirky%20Female%20English_OneShot.MP3" controls></audio>
Better reproduces the<br>
exaggerated high pitch<br>
characteristic of anime voices<br>
</td>
<td>
<audio class="audio-sm" src="assets/audios/ElevenLabs_Quirky%20Female%20English.mp3" controls></audio>
</td>
</tr>
<tr>
<td>
<audio class="audio-sm" src="assets/audios/Neurotic%20Teenage%20English_Prompt.MP3" controls></audio>
</td>
<td>
Oh my gosh, like, I literally can't believe<br>
what just happened! Um, so basically, I was,<br>
you know, just sitting there in class,<br>
right? And then, ugh, this totally weird<br>
thing happened - like, seriously weird! Wait,<br>
wait... Should I even be talking about this?<br>
Ugh, whatever.
</td>
<td>
<audio class="audio-sm" src="assets/audios/Neurotic%20Teenage%20English_ZeroShot.MP3"
controls></audio>
Effectively follows textual cues<br>
for both longer and shorter<br>
inter-sentence pauses
</td>
<td>
<audio class="audio-sm" src="assets/audios/Neurotic%20Teenage%20English_OneShot.MP3" controls></audio>
Better reproduces the<br>
exaggerated high pitch<br>
characteristic of anime voices<br>
</td>
<td>
<audio class="audio-sm" src="assets/audios/ElevenLabs_Neurotic%20Teenage%20English.mp3"
controls></audio>
</td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="article-block">
<h2 id="multilingual-and-cross-lingual-capabilities-demonstrations">Multilingual and Cross-Lingual Capabilities
Demonstrations</h2>
<p>Speech-02-HD maintains high naturalness in less common languages while demonstrating significant advantages
in
Standard
Chinese pronunciation accuracy.</p>
<div class="scroll-wrapper" style="margin-top: 2rem;">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col">Languages</th>
<th scope="col">Source Audio</th>
<th scope="col">Text</th>
<th scope="col">MiniMax<br>Speech_02_HD</th>
<th scope="col">ElevenLabs<br>Multilingual_v2</th>
<th scope="col">OpenAI<br>TTS_1_HD<br>(*not cloned voice)</th>
</tr>
<!-- Thai -->
<tr class="border-bottom-thin">
<th>Thai</th>
<td>
<audio class="audio-sm" src="assets/audios/Thai_Male_Sourse.wav" controls></audio>
</td>
<td>
สวัสดีค่ะ วันนี้อากาศดีมากเลย<br>
คุณจะไปทานอาหารกลางวันที่ไหนคะ<br>
ฉันกำลังคิดว่าจะไปร้านอาหารไทยแถวนี้<br>
</td>
<td>
<audio class="audio-sm" src="assets/audios/Thai.mp3" controls></audio>
</td>
<td>
Thai not perfectly supported
<audio class="audio-sm" src="assets/audios/ElevenLabs_Thai.mp3" controls></audio>
</td>
<td>
<audio class="audio-sm" src="assets/audios/OpenAI_Thai.mp3" controls></audio>
</td>
</tr>
<!-- Vietnamese -->
<tr class="border-bottom-thin">
<th>Vietnamese</th>
<td>
<audio class="audio-sm" src="assets/audios/Vietnamese_Female_Sourse.wav" controls></audio>
</td>
<td>
Tôi đang đọc một cuốn sách rất hay về lịch sử Việt Nam.<br>
Những câu chuyện về văn hóa truyền<br>
thống thật sự rất thú vị.<br>
</td>
<td>
<audio class="audio-sm" src="assets/audios/Vietnamese.mp3" controls></audio>
</td>
<td>
Vietnamese not perfectly supported
<audio class="audio-sm" src="assets/audios/ElevenLabs_Vietnamese.mp3" controls></audio>
</td>
<td>
<audio class="audio-sm" src="assets/audios/OpenAI_Vietnamese.mp3" controls></audio>
</td>
</tr>
<!-- Czech -->
<tr class="border-bottom-thin">
<th>Czech</th>
<td>
<audio class="audio-sm" src="assets/audios/Czech_Female_Sourse.wav" controls></audio>
</td>
<td>
Ranní mlha se pomalu zvedá nad řekou,<br>
zatímco první paprsky slunce prosvítají mezi stromy.<br>
Ptáci začínají svůj ranní koncert.<br>
</td>
<td>
<audio class="audio-sm" src="assets/audios/Czech.mp3" controls></audio>
</td>
<td>
<audio class="audio-sm" src="assets/audios/ElevenLabs_Czech.mp3" controls></audio>
</td>
<td>
<audio class="audio-sm" src="assets/audios/OpenAI_Czech.mp3" controls></audio>
</td>
</tr>
<!-- Polish -->
<tr class="border-bottom-thin">
<th>Polish</th>
<td>
<audio class="audio-sm" src="assets/audios/Polish_Male_Sourse.wav" controls></audio>、
</td>
<td>
Młoda sowa siedzi cicho na gałęzi sosny,<br>
obserwując leśną polanę w świetle księżyca.<br>
Wiatr delikatnie porusza liśćmi drzew.<br>
</td>
<td>
<audio class="audio-sm" src="assets/audios/Polish.mp3" controls></audio>
</td>
<td>
<audio class="audio-sm" src="assets/audios/ElevenLabs_Polish.mp3" controls></audio>
</td>
<td>
<audio class="audio-sm" src="assets/audios/OpenAI_Polish.mp3" controls></audio>
</td>
</tr>
<!-- Japanese -->
<tr class="border-bottom-thin">
<th>Japanese</th>
<td>
<audio class="audio-sm" src="assets/audios/Japanese_DominantMan_Sourse.mp3" controls></audio>
</td>
<td>
電車が遅延している影響で、渋谷駅がとても混雑<br>
しています。次の山手線は約10分後に到着<br>
予定です。お急ぎのお客様は、他の路線も<br>
ご利用ください。
</td>
<td>
<audio class="audio-sm" src="assets/audios/Japanese.mp3" controls></audio>
</td>
<td>
<audio class="audio-sm" src="assets/audios/ElevenLabs_Japanese_Dominant_Man.mp3" controls></audio>
</td>
<td>
<audio class="audio-sm" src="assets/audios/OpenAI_Japanese.mp3" controls></audio>
</td>
</tr>
</tbody>
</table>
</div>
<p style="margin-top: 4rem;">Speech-02-HD has superior performance in zero-shot cross-lingual scenarios.</p>
<div class="scroll-wrapper" style="margin-top: 2rem;">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col">Original Language</th>
<th scope="col">Source Audio</th>
<th scope="col">Mixed Language</th>
<th scope="col">Text</th>
<th scope="col">MiniMax<br>Speech_02_HD</th>
<th scope="col">ElevenLabs<br>Multilingual_v2</th>
<th scope="col">OpenAI<br>TTS_1_HD<br>(*not cloned voice)</th>
</tr>
<tr class="border-bottom-thin">
<td>English</td>
<td>
<audio class="audio-sm" src="assets/audios/Wong_Sourse.mp3" controls></audio>
</td>
<td>English + Mandarin</td>
<td>
Kiddo! Come come come, 学如逆水行舟，不进则退。<br>
I see you're using AI tools already - so smart!<br>
But eh, cannot just rely on tools only lah!<br>
The future belongs to those who can work alongside AI,<br>
not those scared of it.
</td>
<td>
<audio class="audio-md" src="assets/audios/English-Mandarin.wav" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/ElevenLabs_English-Mandarin.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/OpenAI_English-Mandarin.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>Mandarin</td>
<td>
<audio class="audio-sm" src="assets/audios/ShiBanYu_Sourse.mp3" controls></audio>
</td>
<td>Mandarin + Cantonese</td>
<td>
老铁啊，多谢晒你送我呢本，广州话正音字典，咁好嘢喎！<br>
我呢个大老爷们儿学广州话真系好难㗎！成日都分唔清声调啊。<br>
嗱，而家有咗呢本书，什么都好啦。
</td>
<td>
<audio class="audio-md" src="assets/audios/Mandarin-Cantonese.MP3" controls></audio>
</td>
<td>
Cantonese not supported
</td>
<td>
Cantonese not supported
</td>
</tr>
<tr class="border-bottom-thin">
<td>Mandarin</td>
<td>
<audio class="audio-sm" src="assets/audios/ShuanQ_Sourse.mp3" controls></audio>
</td>
<td>Mandarin + English</td>
<td>
The people said, 桂林's scenery is the first under heaven.<br>
Yet in my opinion, 阳朔 scenery is better than 桂林。<br>
群峰倒影山浮水，无水无山不入神。
</td>
<td>
<audio class="audio-md" src="assets/audios/Mandarin-English.WAV" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/ElevenLabs_Mandarin-English.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/OpenAI_Mandarin-English.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>English</td>
<td>
<audio class="audio-sm" src="assets/audios/CoCo_Sourse.mp3" controls></audio>
</td>
<td>English + Spanish</td>
<td>
Mi abuelita always told me "el que persevera, alcanza".<br>
If you persevere, you'll achieve your dreams!<br>
Guess what! They choose me to play the lead role in our BIG show!
</td>
<td>
<audio class="audio-md" src="assets/audios/English-Spanish.wav" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/ElevenLabs_English-Spanish.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/OpenAI_English-Spanish.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>Japanese</td>
<td>
<audio class="audio-sm" src="assets/audios/Powerful_Girl_Sourse.mp3" controls></audio>
</td>
<td>Japanese + Korean</td>
<td>
最近の天気予報によりますと、今週末は桜の開花に最適<br>
な気温になる予定です。<br>
東京都内の各公園では花見客で賑わうことが予想されますが、<br>
서울에서도 벚꽃이 피기 시작했다고 하네요.<br>
이번 주말에는 여의도 공원에서 벚꽃 축제가 열린다고 하니<br>
많은 분들이 찾아오실 것 같습니다.
</td>
<td>
<audio class="audio-md" src="assets/audios/Japanese-Korean.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/ElevenLabs_Japanese-Korean.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/OpenAI_Japanese-Korean.mp3" controls></audio>
</td>
</tr>
</tbody>
</table>
</div>
<p>*Although OpenAI currently does not support voice cloning functionality, we still wish to conduct comparative
listening
tests with its excellent naturalness as a reference.</p>
</div>
<div class="article-block">
<h2 id="flow-vae-vs-vae-comparisons">Flow-VAE vs. VAE Comparison</h2>
<p>Flow-VAE is less likely to produce the following instabilities.</p>
<div class="scroll-wrapper" style="margin-top: 2rem;">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col" style="text-align: center;">Source Audio</th>
<th scope="col" style="text-align: center;">Flow-VAE</th>
<th scope="col" style="text-align: center;">VAE</th>
<th scope="col" style="text-align: center;">Differences</th>
</tr>
<tr class="border-bottom-thin">
<td style="width: 25%">
<audio src="assets/audios/Condition1.wav" controls></audio>
</td>
<td style="width: 25%">
<audio src="assets/audios/FlowVAE1.wav" controls></audio>
</td>
<td style="width: 25%">
<audio src="assets/audios/VAE1.wav" controls></audio>
</td>
<td>
Flow-VAE reproduces more continuous<br>
and natural reverberation
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio src="assets/audios/Condition2.wav" controls></audio>
</td>
<td>
<audio src="assets/audios/FlowVAE2.wav" controls></audio>
</td>
<td>
<audio src="assets/audios/VAE2.wav" controls></audio>
</td>
<td>
VAE introduces unwanted<br>
high-frequency components
</td>
</tr>
<tr>
<td>
<audio src="assets/audios/Conditon3.wav" controls></audio>
</td>
<td>
<audio src="assets/audios/FlowVAE3.wav" controls></audio>
</td>
<td>
<audio src="assets/audios/VAE3.wav" controls></audio>
</td>
<td>
VAE produces electronic-sounding<br>
artifacts at the beginning
</td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="article-block">
<h2 id="professional-voice-clone-pvc-demonstrations">Professional Voice Clone (PVC) Demonstrations</h2>
<p>For more complex dialectal accents and tonal characteristics, PVC can reproduce these features while
maintaining high
naturalness based on the text content.</p>
<div class="scroll-wrapper" style="margin-top: 2rem;">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col" style="text-align: center;">Source Audio</th>
<th scope="col" style="text-align: center;">Zero-Shot</th>
<th scope="col" style="text-align: center;">PVC</th>
<th scope="col" style="text-align: center;">Differences</th>
</tr>
<tr class="border-bottom-thin">
<td style="width: 25%">
<audio src="assets/audios/JosephBrodsky_Source.wav" controls></audio>
</td>
<td style="width: 25%">
<audio src="assets/audios/JosephBrodsky_Fast.mp3" controls></audio>
</td>
<td style="width: 25%">
<audio src="assets/audios/JosephBrodsky_PVC.mp3" controls></audio>
</td>
<td>
Like the ZeroShot version, the PVC<br>
version has rising sentence-final intonation,<br>
but distinctively sustains this<br>
elevated pitch instead of the typical<br>
pitch declination found in common<br>
declarative sentences
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio src="assets/audios/TianJin_Source.wav" controls></audio>
</td>
<td>
<audio src="assets/audios/TianJin_Fast.mp3" controls></audio>
</td>
<td>
<audio src="assets/audios/TianJin_PVC.mp3" controls></audio>
</td>
<td>
With more materials, the model not only<br>
reproduces the speaker's voice characteristics<br>
but also accurately captures more<br>
dialectal features
</td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="article-block">
<h2 id="emotion-control-demonstrations">Emotion Control Demonstrations</h2>
<h3>Source Audio for Refreshing Young Man</h3>
<audio src="assets/audios/Mandarin_Refreshing_Young_Man_Sourse.mp3" controls></audio>
<h3>DEMO</h3>
<div class="scroll-wrapper">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col">Neutral</th>
<th scope="col" style="min-width: 120px;">Emotion</th>
<th scope="col">Text</th>
<th scope="col">Emotion Control Audio</th>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-md" src="assets/audios/Neutral1.mp3" controls></audio>
</td>
<td>
Surprised
</td>
<td>
天哪！我完全没想到会在这里遇见你，<br>
都过去这么多年了，你一点都没变！
</td>
<td>
<audio class="audio-md" src="assets/audios/Surprised.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-md" src="assets/audios/Neutral2.mp3" controls></audio>
</td>
<td>
Disgusted
</td>
<td>
这个地方实在太脏乱了，到处都是垃圾和难闻的气味儿，<br>
我一秒钟都不想多待。
</td>
<td>
<audio class="audio-md" src="assets/audios/Disgusted.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-md" src="assets/audios/Neutral3.mp3" controls></audio>
</td>
<td>
Fearful
</td>
<td>
深夜回家的路上，我清楚地听见身后有脚步声在跟着我，<br>
可是回头却什么都看不见。
</td>
<td>
<audio class="audio-md" src="assets/audios/Fearful.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-md" src="assets/audios/Neutral4.mp3" controls></audio>
</td>
<td>
Angry
</td>
<td>
我付出了这么多，换来的却是这样的背叛！<br>
你怎么可以这样对待我的信任！
</td>
<td>
<audio class="audio-md" src="assets/audios/Angry.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-md" src="assets/audios/Neutral5.mp3" controls></audio>
</td>
<td>
Sad
</td>
<td>
躺在床上翻来覆去，心里压着说不出的难过和沮丧，<br>
昨天晚上又失眠了。
</td>
<td>
<audio class="audio-md" src="assets/audios/Sad.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
<audio class="audio-md" src="assets/audios/Neutral6.mp3" controls></audio>
</td>
<td>
Happy
</td>
<td>
和好朋友一起在院子里烧烤，聊着有趣的故事，<br>
享受着美食和欢乐的时光。
</td>
<td>
<audio class="audio-md" src="assets/audios/Happy.mp3" controls></audio>
</td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="article-block">
<h2 id="text-prompted-voice-generation-demonstrations">Text-Prompted Voice Generation Demonstrations</h2>
<div class="scroll-wrapper">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col">Prompt</th>
<th scope="col">Text</th>
<th scope="col" style="text-align: center;">Audio</th>
</tr>
<tr class="border-bottom-thin">
<td>
男性中年声音，说中文，音色浑厚醇厚，带有自然的磁性，<br>
语速偏慢，音量适中，音调偏低沉。声音整体给人沉稳可靠的感觉，<br>
在深度访谈场景中表现出专业性和亲和力，音质清晰，吐字规整有力。
</td>
<td>
在这个安静的夜晚，让我们一起走进《人生笔记》这本书。<br>
作者用平实的文字记录下生活中的点点滴滴，<br>
让我们看到平凡中的真善美。<br>
今天，我们先来读第一章：'生活的痕迹'......
</td>
<td>
<audio class="audio-md" src="assets/audios/深度访谈男中年.wav" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
说中文的女青年，音色偏甜美，语速比较快，<br>
说话时带着一种轻快的感觉，整体音调较高，像是在直播带货，<br>
整体氛围比较活跃，声音清晰，听起来很有亲和力。
</td>
<td>
亲爱的宝宝们，等了好久的神仙面霜终于到货啦！<br>
你们看这个包装是不是超级精致？<br>
我自己已经用了一个月了，效果真的绝绝子！<br>
而且这次活动价真的太划算了，错过真的会后悔的哦～
</td>
<td>
<audio class="audio-md" src="assets/audios/直播带货女青年.wav" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
中国男性声音，听着像是青年，音色清亮，语速比较快，<br>
说话很有激情，像是在解说比赛，声音中带着紧张和兴奋的感觉。
</td>
<td>
漂亮！这个进攻太精彩了！张伟突破防线，<br>
一个漂亮的转身，球传到禁区，王超跟上，射门！<br>
球进了！难以置信的精彩配合，现场观众都沸腾了！
</td>
<td>
<audio class="audio-md" src="assets/audios/体育解说男青年.wav" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
中国女青年的声音，音色清脆，说话速度偏快，语调活泼，<br>
像是在做游戏直播，声音中带着愉快的感觉，整体音调较高，<br>
整体氛围比较轻松。
</td>
<td>
啊！这里有个宝箱！让我们看看里面是什么～<br>
哇！是传说中的紫色装备！运气也太好了吧！<br>
谢谢小伙伴们的打赏，我们继续往前探索......
</td>
<td>
<audio class="audio-md" src="assets/audios/游戏主播女青年.wav" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
English-speaking female voice, sounding relatively young,<br>
with a sweet and pleasant tone. Speaking at a moderate pace<br>
with a touch of energy, similar to someone narrating a<br>
beauty/makeup tutorial video. The overall atmosphere is<br>
relaxed and cheerful.
</td>
<td>
Hi everyone! Today I'll be sharing a soft, romantic<br>
makeup look that's perfect for dates. Many of you have <br>
been asking how to apply this eyeshadow naturally - the<br>
key is using gentle techniques. Let's go through the<br>
steps together...
</td>
<td>
<audio class="audio-md" src="assets/audios/美妆女博主.wav" controls></audio>
</td>
</tr>
<tr>
<td>
English-speaking middle-aged male voice, slightly husky, <br>
speaking at a moderate-to-slow pace with a deep tone. Like<br>
someone telling an old story, conveying a nostalgic feeling,<br>
with a relaxed and composed manner of speaking.
</td>
<td>
That was back in the late 1970s. I remember when our <br>
village first got electricity - everyone was so excited. <br>
In theevenings, people would bring their stools and <br>
gather under the big banyan tree by the village committee <br>
office to watch movies projected on the wall. Even now, <br>
thinking back to those moments still fills me with warmth.
</td>
<td>
<audio class="audio-md" src="assets/audios/回忆男中年.wav" controls></audio>
</td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="article-block">
<h2 id="comparison-of-voice-naturalness">Comparison of voice naturalness
with the previous generation products</h2>
<p>The new model demonstrates significant advantages in naturalness compared to the previous version.</p>
<h3 style="margin-top: 2rem;">Source Audio for Radiant_Girl</h3>
<audio src="assets/audios/English_Radiant_Girl_Sourse.wav" controls></audio>
<h3>DEMO</h3>
<div class="scroll-wrapper">
<table style="width: 100%;">
<tbody>
<tr class="border-bottom-thin">
<th scope="col">Text</th>
<th scope="col" style="text-align: center;">MiniMax<br>Speech_02_HD</th>
<th scope="col" style="text-align: center;">Microsoft<br>Azure TTS</th>
<th scope="col" style="text-align: center;">AWS<br>Polly</th>
</tr>
<tr class="border-bottom-thin">
<td>
I sat alone in the empty room, staring at the old photographs,<br>
wondering how everything could change so quickly,<br>
how a lifetime of memories could fade away just like that.
</td>
<td>
<audio class="audio-md" src="assets/audios/Radiant_Girl_1.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Emma_1.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Joanna_1.mp3" controls></audio>
</td>
</tr>
<tr class="border-bottom-thin">
<td>
The moment I held my acceptance letter, my heart burst with joy - <br>
all those sleepless nights finally paid off, and I couldn't stop<br>
dancing around the room, calling everyone I knew to share this amazing news!
</td>
<td>
<audio class="audio-md" src="assets/audios/Radiant_Girl_2.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Emma_2.mp3" controls></audio>
</td>
<td>
<audio class="audio-md" src="assets/audios/Joanna_2.mp3" controls></audio>
</td>
</tr>
</tbody>
</table>
</div>
</div>
<div class="article-block">
<h2 id="citation">Citation</h2>
<div>
<pre>
<code>
@misc{minimax2025minimaxspeechintrinsiczeroshottexttospeech,
title={MiniMax-Speech: Intrinsic Zero-Shot Text-to-Speech with a Learnable Speaker Encoder},
author={Bowen Zhang, Congchao Guo, Geng Yang, Hang Yu, Haozhe Zhang, Heidi Lei, Jialong Mai, Junjie Yan, Kaiyue Yang,
Mingqi Yang, Peikai Huang, Ruiyang Jin, Sitan Jiang, Weihua Cheng, Yawei Li, Yichen Xiao, Yiying Zhou, Yongmao Zhang,
Yuan Lu, Yucen He},
year={2025},
eprint={2505.07916},
archivePrefix={arXiv},
primaryClass={eess.AS},
url={https://arxiv.org/abs/2505.07916},
}</code>
</pre>
</div>
</div>
</article>
</main>
<script>
MathJax = {
tex: {
inlineMath: [['$', '$'],],
},
}
const darkModeToggle = document.getElementById('dark-mode-toggle')
darkModeToggle.addEventListener('click', () => {
document.body.classList.toggle('latex-dark')
})
</script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js"></script>
</body>
</html>