Controllable generation
Target text, a natural-language style instruction, and reference audio are composed into expressive speech.
- Speaker and acoustic attributes
- Detailed speaking-style descriptions
- Role-playing scenarios
Bilingual speech intelligence
Voice follows language.
Language learns from voice.
Cyclespeech brings controllable speech generation and structured speech understanding into one bilingual framework.
Paper, code, and model release details coming soon.
Cycle
shared speech intelligence
Generatetext → speech
Understandspeech → structure
2complementary modes
EN · ZHbilingual inference
APS · DSD · RPinstruction views
One model, two directions
A speech system that can both perform and perceive.
Describe how a line should sound, then synthesize it with a reference voice. Or provide speech and recover a structured account of what was said and how it was delivered.
Structured understanding
Speech is analyzed into a consistent voice-profile JSON object together with its transcript.
- Language and speaker attributes
- Emotion, scene, and delivery
- Complete speech transcription
Audio showcase
From direction to delivery.
Compare reference speech with outputs from Step-Audio-2-mini, Cyclespeech(SFT), and Cyclespeech(CycleGRPO) under identical inputs across all three instruction types. Headphones recommended.
ENAPS
01
Target text
“So, I wanted to create a dish which invoked that sensation”
Speaker & acoustic instruction
gender: Male; age: young adult; timbre / pitch: baritone, smooth, resonant, polished, warm, velvety, breathy; speech rate: moderate, rhythmic; volume: normal; accent annotation: British, contemporary Southern English (Modern Received Pronunciation).
ZHAPS
02
目标文本
“那也不要扫孤王的兴嘛”
说话人与声学属性指令
男性，青壮年；中低音，浑厚磁性、细腻圆润；语速较快，音量适中；标准普通话。
ENDSD
03
Target text
“does the invitation for me to stay still stand”
Detailed style description
Speak in a General American accent slowly, in a deep baritone register, with velvety and resonant timbre, at soft volume. Keep the tone vulnerable and tentative, with an emotionally presence, letting the delivery sound sensitive.
ZHDSD
04
目标文本
“你说咱们一家还愁避难的地方吗？”
详细风格描述
展现出中高音声线与清脆、明亮、富有弹性、金属般清透的音色质感，表现出说话人强烈的自豪感和松弛感，并夹杂自信，展现出冷静、自信的性格特点，语音韵律节奏感强，语速中等偏快。
ENRP
05
Target text
“Maybe Ty will get the hint if he sees me in this on my Instagram”
Role-playing instruction
In a cozy bedroom or dressing room, a young woman in her late teens or early twenties stands before a mirror or phone camera, her eyes sparkling with mischief.
ZHRP
06
目标文本
“我要让他截肢，他硬是不肯。”
角色扮演指令
在医院的走廊里，一名年轻女性紧握双手，声音颤抖地向医生诉说。
Application · Video dubbing
Speech that belongs in the scene.
These examples pair Cyclespeech speech generation with video to illustrate a dubbing workflow. The model uses a reference voice, target text, and a style instruction.
EN
01
Target line
“Oh my God! Do you remember that bottle of wine we put aside the night Haley was born?”
Speaking instruction
In a car, an adult woman suddenly remembers a family keepsake while speaking to her husband. She reacts with an immediate, spontaneous burst of surprise, then asks the question at a brisk, lively conversational pace with minimal pauses.
中文
02
目标台词
“在座的各位，你们都肩负着，光明县，经济、社会、持续健康发展重任。”
表演指令
一位男性领导在进行正式讲话，语调沉稳有力，语速缓慢，旨在强调听众肩负的重要职责，整体情感庄重而富有号召力。
Structured listening
Speech becomes language you can inspect.
The understanding mode maps an input waveform to a stable voice-profile object. It captures the transcript alongside speaker, delivery, emotion, and scene attributes.
genderageaccent
emotionemotion_detailpersonality
scenespeedvolumetexture
transcript
See the shared framework
{
"gender": "...",
"age": "...",
"accent": "...",
"emotion": "...",
"emotion_detail": ["..."],
"personality": "...",
"scene": "...",
"speed": "...",
"volume": "...",
"texture": "...",
"txt": "full transcript"
}Understanding examples
Listen to the input. Inspect the prediction.
Each example pairs an unseen test utterance with Cyclespeech's structured analysis.
ENInput speech
01
Transcript
“Come on, Kong! Let's make like a banana and split.”
Cyclespeech prediction
SpeakerYoung woman · General American
AffectHappy · playful urgency
DeliveryMedium-fast · normal volume
TextureBright, crisp, and resonant
ContextLively outdoor setting
PersonalityEnergetic, confident, and engaging
中文输入语音
02
转写文本
“我怎么就泼皮无赖了？”
Cyclespeech 预测
说话人青年男性 · 普通话
情感生气 · 强烈辩解
表达快速 · 大音量
音色清脆但紧绷
场景冲突场景
人物性格冲动、委屈
ENInput speech
03
Transcript
“They tell you everything they do is for the family.”
Cyclespeech prediction
SpeakerMiddle-aged woman · General American
AffectSad · somber resignation
DeliveryMedium-slow · normal volume
TextureVelvety, smoky, and breathy
ContextDimly lit room
PersonalityReflective, weary, and sincere
中文输入语音
04
转写文本
“因为谁都搬不动你。”
Cyclespeech 预测
说话人青年女性 · 普通话
情感开心 · 轻松调侃
表达中快语速 · 正常音量
音色清脆甜美
场景轻松的社交场合
人物性格亲和力强
The cycle
Complementary tasks, shared speech intelligence.
Cyclespeech connects generation and understanding around a shared audio-language foundation, with task-specific adapters for each direction.
Style instruction
Target text
Reference audio
generation
shared foundationCyclespeechswitchable task adapters
speech
audio input
Waveform
Voice profile
Transcript
Explore the project
Listen, generate, and understand.
Explore bilingual generation and structured understanding examples.