Speech Synthesis TRIZ Cases
- ACELP Speech Codebook Optimization Using Toeplitz Autocorrelation
A Toeplitz autocorrelation matrix cuts ACELP codebook search complexity and memory use while preserving speech perceptual quality.
- MFCC Frequency Shifting for Clearer Speech Perception
Maps high-frequency speech power to lower mel bands and resynthesizes MFCC-based audio to improve clarity for hearing-impaired listeners.
- AI Speech Synthesis Pipeline for Low-Latency Prosody Generation
Sequential phonetic and prosodic generation with an LM-TTS adaptor cuts speech latency while improving accuracy, style, and emotion.
- Text-to-Speech Acoustic Control Without Retraining
A two-stage TTS workflow uses control vectors to adjust pitch, energy, and duration without extra training while preserving speech quality.
- Speech Synthesis Model Training With Targeted Text Feedback
Performance metrics flag weak text sequences so speech models can request targeted actor recordings and cut training time and data use.
- Text-to-Speech Adaptation Using Cursor-Tracked Reading Pace
Cursor position and motion guide TTS pace and speech characteristics, improving reading alignment, engagement, and comprehension.
- Speech Synthesis Acoustic Feature Adjustment for Pronunciation and Prosody
Adjustment dictionaries target repeated pronunciation and prosody errors in neural speech synthesis without repeated user input.
- Expressive Neural TTS Training Using Expressivity-Scored Data
A two-phase, expressivity-scored training approach helps neural TTS generate more natural speech with stronger emotional expression.
- Audio Transformation Model for Clean Speech From Noisy Recordings
Machine learning extracts noise and speech features to synthesize clean voice audio from mobile or non-studio recordings.
- Neural Spline Normalizing Flows for Natural Speech Synthesis
Neural spline flows model pitch, energy, and unvoiced regions to make text-to-speech output more natural and expressive.
- Phoneme Pitch Adjustment Using Binned Vectors for Expressive Speech
Binned phoneme pitch vectors and neural prediction layers enable fine pitch tuning for more expressive and context-aware text-to-speech.
- Variable-Rate TTS Key Frames for Low-Latency Speech Synthesis
A neural TTS model skips less informative acoustic frames and interpolates them later to cut latency and bandwidth without degrading speech quality.
- Text-to-Speech Attention Alignment for Natural Timing
Thresholded and activated attention vectors improve text-to-speech alignment, yielding more natural timing and better non-speech synthesis with limited data.
- Synthetic Speaker Embeddings for Natural and Private Speech
Multi-stage embedding training enables natural synthetic voices while reducing traceability to real speakers and resisting noise defects.
- Text-Conditional Audio Generation for DAS Domain Shift
A two-stage audio pipeline filters hallucinated events and shifts signals with impulse response and noise to build DAS-ready datasets.
- Multilingual Speech Synthesis Using Disentangled Speaker-Language Features
By removing speaker and language cues during training and reapplying them at inference, this case improves multilingual speech quality without fine-tuning.
- LLM Audio Generation Using Role and Emotion Reference Matching
Role and emotion extraction guides reference text and audio selection, improving synthesis authenticity while reducing labeled-data training cost.
- Synthetic Voice Parameter Search for Minimal Recording Input
A 2D voice search and parameter-mixing approach builds a target-like synthetic voice when only minimal recordings are available.
- Prosody Prediction for Expressive Text-to-Speech Synthesis
A shared linguistic-acoustic prosody model adds contextual and speaker cues to generate more natural, engaging synthesized speech.
- Neural Vocoder Subband Generation for Real-Time Speech Synthesis
Simultaneous subband generation cuts neural vocoder forward propagations, enabling faster real-time speech waveform synthesis.
- Emotion Speech Synthesis Using Decoupled Timbre and Emotion Mapping
Separating timbre and emotion features cuts training data needs while preserving consistent voice quality and emotion intensity in synthesized speech.
- Real-Time Accent Mimicking That Preserves Natural Voice
Machine learning extracts target accent features and adapts live speech while preserving the speaker's natural voice for clearer conversations.