What's in a Voice? Deconstructing the Elements of Realistic Speech Synthesis

Introduction: Beyond Robotic Monotones – The New Frontier of AI Speech

The evolution of Text-to-Speech (TTS) technology has been a remarkable journey. We've moved from the stilted, robotic monotones of early computer systems to the remarkably fluid and intelligible voices that power today's digital assistants and applications.¹ Yet, as we've closed the gap on basic pronunciation and clarity, a new, more subtle challenge has emerged: the uncanny valley of voice. This is the realm of AI speech that is technically correct but emotionally vacant, grammatically perfect but conversationally awkward. It sounds almost human, but the absence of soul, warmth, and the beautiful imperfections of genuine speech creates a jarring, sterile experience.

True vocal realism is not merely about converting text into sound. It is a complex symphony of nuanced elements that must be meticulously engineered and harmonized. This report deconstructs the anatomy of a truly realistic AI voice, exploring the "what" and the "how" behind the three pillars of natural speech:

• Prosody: The fundamental music and rhythm of speech that conveys meaning beyond words
• Emotional Inflection: The rich palette of feelings and attitudes that colors our communication
• Disfluencies: The authentic hesitations, pauses, and filler words that signal a thinking, breathing human at the other end

For developers, content creators, and businesses, understanding these components is crucial for creating engaging, immersive, and believable audio experiences. This guide will delve into the linguistic principles and the sophisticated machine learning models that bring these elements to life. Furthermore, it will navigate the competitive landscape of TTS solutions, highlighting a critical issue that often hinders creative workflows: the digital artifact problem. As we will explore, next-generation platforms like VocalCopycat are leading the charge not just by mimicking the nuances of human speech, but by delivering them with a pristine, artifact-free quality that finally empowers creators to focus on their vision, not on tedious audio correction.³

Section 1: The Soul of Speech – Mastering Prosody and Intonation

Defining Prosody: The Music Behind the Words

At the heart of natural-sounding speech lies prosody, a term that describes the tune, rhythm, and melody of language.⁵ It is a foundational layer of meaning that operates above the level of individual sounds (phonemes), often spanning entire phrases and sentences. Linguists refer to these features as "suprasegmentals" because they are layered over the basic segments of speech, providing context, emphasis, and structure.⁶

The core acoustic building blocks of prosody are:

• Vocal Pitch: The perceived highness or lowness of the voice, measured physically as the fundamental frequency (F0)
• Loudness: The perceived volume of the voice, measured as acoustic intensity
• Rhythm: The pattern of timing and duration given to phonemes and syllables, creating the cadence of speech⁶

These three elements are the raw materials that AI developers must learn to model and control to create a voice that sounds less like a machine reading words and more like a human communicating ideas. The challenge is compounded by the fact that these same acoustic features are also the primary carriers of other layers of information. Beyond the direct linguistic meaning of words, prosody conveys paralinguistic information, such as a speaker's attitude (irony, sarcasm), and non-linguistic information, like a speaker's emotional state, health, or even their membership in a particular speech community.⁶

This profound overlap is a major source of complexity in speech synthesis. A simple rise in pitch, for instance, could signify a question (a linguistic function), a state of excitement (an emotional function), or a sarcastic tone (a paralinguistic function). A truly advanced TTS system must therefore be acutely context-aware. It cannot rely on simple, rule-based models that map punctuation to a specific prosodic change. Instead, it must develop a deeper understanding of grammar, semantics, and pragmatics to disentangle these intertwined signals and generate an output that is not just acoustically correct, but meaningfully appropriate. This is the fundamental reason why early, rule-based TTS systems sounded so unnatural and why modern neural network-based approaches are achieving unprecedented levels of realism.

The Power of Pitch: Intonation Contours

Of all the prosodic elements, intonation—the variation of pitch over time—is perhaps the most powerful conveyor of meaning. It is the melody of a sentence, capable of transforming a simple statement into a question with a single upward inflection.⁷ For example, the words "See you tomorrow" can be a declarative statement with a falling pitch or an inquisitive question with a rising pitch, all without changing a single word.⁷

From a technical standpoint, modeling intonation involves generating a smooth and natural fundamental frequency (F0) contour that maps onto the text. Early synthesis programs struggled with this, often producing unnatural-sounding speech by placing an F0 peak on every lexical word, a pattern rarely found in human speech.⁸ More sophisticated models have learned that pitch peaks should align with stressed syllables and that the overall F0 contour of a declarative sentence typically exhibits a gradual downward slope, a phenomenon known as declination. The most advanced models can even account for this declination implicitly, understanding that it forms a baseline against which prominence is measured, even if the surface-level pitch contour doesn't show a steady decline.⁸

A significant practical challenge, particularly for developers of real-time conversational agents and voice bots, is the "end-of-sentence" problem. Many TTS systems are designed to treat any submitted text chunk as a complete sentence, automatically applying a falling, declarative intonation at the end.⁹ When synthesizing partial sentences streamed from a Large Language Model (LLM), this results in a choppy, unnatural conversational flow where the AI sounds like it's constantly finishing its thought, even mid-sentence. This highlights the critical need for TTS models that can process partial utterances and intelligently predict a continuing, rather than a terminal, intonation contour.

The Developer's Toolkit for Prosody Control

To give developers and creators fine-grained control over these prosodic elements, the industry relies on a standardized tool: the Speech Synthesis Markup Language (SSML). SSML is an XML-based language that allows one to embed explicit instructions within a text input, guiding the synthesis engine on how to render the speech.⁹

The primary SSML tags for managing prosody are:

• The

  <prosody>

  Tag: This is the workhorse for prosodic modification, enabling direct control over pitch, rate (speaking speed), and volume (loudness).⁹ For instance, a developer could make a voice sound more excited by increasing its pitch and rate using:

  <prosody pitch="high" rate="fast">This is exciting!</prosody>
  ```<sup>13</sup>
  
  

• The

  <break>

  Tag: This tag is essential for controlling rhythm and pacing by inserting pauses. Pauses can be defined by their relative strength—such as weak for the pause after a comma or strong for the pause after a period—or by a precise duration in seconds or milliseconds.⁹ This control is vital for preventing speech from sounding rushed and for giving listeners time to process information.

• The

  <emphasis>

  Tag: This allows for adding stress to specific words or phrases, which is typically achieved by the model through a combination of increased pitch, loudness, and duration.⁹

The very existence of SSML underscores the limitations of earlier TTS engines; they were passive renderers that needed to be explicitly told how to speak. While SSML remains an indispensable tool for artistic control and specific use cases, the trajectory of modern TTS is toward greater automation and intelligence. The most advanced platforms now perform sophisticated linguistic analysis on the input text, using part-of-speech tags, syntactic structure, and punctuation to automatically infer the most natural-sounding prosody.¹¹ A question mark in the text will automatically trigger a rising intonation, and a comma will prompt a slight pause, all without requiring manual SSML tagging.¹³ This creates a powerful, two-tiered system for creators: rely on the model's intelligent, context-aware defaults for speed and convenience, or dive in with SSML for precise, granular control. A superior platform must excel at both.

Section 2: The Color of Conversation – Engineering Emotional Inflection

From Data to Delight (or Despair): The Acoustics of Emotion

Moving beyond correct and natural-sounding prosody, the next frontier in speech synthesis is emotional inflection. Expressive TTS aims to infuse synthesized speech with a rich spectrum of human-like emotional tones, such as happiness, sadness, anger, or excitement.¹⁶ This is achieved by systematically modulating the same acoustic variables that govern prosody. For example, extensive research has shown a strong correlation between a wider pitch range and a higher average pitch with positive, high-energy emotions like joy and excitement. Conversely, a flatter, more monotonous pitch contour is commonly associated with negative emotions like sadness.¹⁸ Similarly, the tempo of speech is a powerful emotional indicator; a fast tempo can signal urgency or enthusiasm, while a slower tempo often reflects thoughtfulness, gravity, or sadness.¹⁸

However, capturing the full complexity of human emotion is one of the most formidable challenges in machine learning.¹⁶ The difficulty arises from the multi-layered nature of vocal expression. As previously noted, a single utterance simultaneously carries linguistic content, paralinguistic cues about attitude, and non-linguistic information about the speaker's emotional state.⁶ A person can be simultaneously sad and ironic, or fearful and sarcastic.⁶ This means that emotional synthesis cannot be a simple, one-to-one mapping where a "sad" tag simply lowers the pitch.

To generate a believable emotional response, a model must understand the broader conversational context. The emotional appropriateness of an utterance is often determined by the dialogue that preceded it. For example, the phrase "You have got to be kidding me" can express genuine surprise, deep frustration, or playful disbelief depending entirely on the context. Advanced research in conversational speech synthesis demonstrates this clearly, showing how the same text can be rendered with a "Surprise" emotion in one conversational flow and a "Disgust" emotion in another, based on the history of the interaction.¹⁹ This necessity for contextual understanding is why early, context-agnostic systems failed to produce believable emotion and why modern, context-aware neural models are beginning to succeed.

Modeling a Spectrum of Feeling: Techniques and Challenges

The foundation of any emotional TTS system is its training data. Models learn to generate emotion by being trained on vast databases of speech that have been carefully annotated with emotional labels.¹⁷ The quality, quantity, and diversity of this data are paramount to the model's success.

Several techniques are used to model and generate these emotions:

• Categorical vs. Dimensional Models: Early systems often relied on a categorical model, treating emotions as discrete, separate labels like "happy," "sad," and "angry".¹⁷ While intuitive, this approach struggles to capture the subtle, blended emotions common in human experience. More advanced systems now employ dimensional models, which plot emotions on a multi-axis space, most commonly using valence (the positivity or negativity of an emotion) and arousal (the intensity or energy level, from calm to excited).¹⁷ This allows for a more continuous and nuanced representation of feeling.

• Style Transfer and Embeddings: A powerful modern approach is to treat emotion as a "style" that can be learned and transferred. In this paradigm, a model can listen to a reference audio clip, extract its emotional style, and apply that style to a completely new sentence.¹¹ This is often accomplished by using a dedicated emotion encoder network. This network analyzes the reference audio and distills its emotional characteristics into a compact mathematical representation known as an embedding. This embedding is then fed as an additional input to the main TTS model, conditioning its output to match the desired emotional style.²³

• Implicit Sentiment from Text: The most sophisticated systems are moving beyond relying solely on audio references or explicit emotion tags. These models leverage powerful language processing models, such as BERT, and advanced architectures like Graph Neural Networks (GNNs) to analyze the input text itself. By examining word choice, sentence structure, and semantic relationships, these models can infer the implicit sentiment of the text, providing a deeper and more authentic emotional foundation for the synthesized speech without any external emotional prompts.²⁴

Despite these advances, two key challenges persist. The first is data sparsity: creating large, high-quality, and accurately labeled emotional speech corpora is incredibly expensive and time-consuming, especially for underrepresented languages and dialects.¹⁶ The second is model complexity: modeling the intricate and dynamic nature of human emotion requires highly sophisticated neural architectures that are computationally intensive to train and deploy.¹⁶

The AI's Emotional Palette: A Look at Commercial Platforms

The leading cloud and AI platforms have adopted various strategies to incorporate emotional expression:

• Microsoft Azure employs deep neural networks where prosody prediction and voice synthesis occur simultaneously, creating a more fluid and natural-sounding output. While they offer extensive prosody control via SSML, their approach to emotion is more reliant on the inherent expressiveness of their neural voices rather than explicit emotion tags.²⁵

• Google Cloud TTS leverages its powerful WaveNet and new Gemini 2.5 models. Their "Chirp 3" voices, based on the AudioLM architecture, are designed for natural intonation.¹⁰ For some specific voices, Google offers explicit emotion control via the SSML tag

  <google:style name="lively">

  .²⁸ The newer Gemini 2.5 models promise even more dynamic performance, with the ability to generate speech with specific emotions and accents upon request.²⁹

• Amazon Polly features several engines, with its "Generative" engine being the most advanced for emotional expression. It can adjust tone and intonation based on the context of the text to produce more engaged and colloquial speech.² However, direct SSML support for a wide range of emotion tags appears to be limited or unsupported in some of their voice engines.³²

• ElevenLabs has established itself as a leader in expressive speech. Their models are designed to interpret emotional context directly from textual cues (e.g., adverbs like "excitedly") and punctuation.³³ Their v3 model is specifically marketed as their "most emotionally rich," and users can influence the output's emotional range and variability using a "Stability" slider, offering a unique blend of automatic interpretation and user-guided control.³³

• Murf.ai and Typecast provide highly accessible, creator-focused solutions. Their platforms feature user-friendly studios where creators can select desired emotions like 'Angry,' 'Cheerful,' or 'Sad' from a simple menu. This offers a more direct and less programmatic method for controlling emotional output, making it ideal for content creators without a deep technical background.³⁶

• VocalCopycat's Approach: Our platform is engineered to capture the natural emotional nuance present in the source text and during the voice cloning process. By building proprietary neural models that are more data-efficient and specifically designed to reduce digital artifacts, we ensure that the generated emotion sounds authentic and clean. This focus on artifact-free quality is a critical differentiator. While some platforms may introduce unwanted noise, distortion, or metallic tones when pushed to generate highly emotional speech, VocalCopycat maintains a consistently believable and polished delivery, ensuring the emotional impact is not undermined by technical glitches.³

Section 3: The Beauty of Imperfection – Why "Um," "Ah," and Pauses Are Crucial

Fluent Disfluency: Redefining "Perfect" Speech

For decades, the holy grail of text-to-speech was perfect, uninterrupted fluency. The goal was to create a voice that never stumbled, paused, or hesitated. However, this pursuit of perfection ironically led to speech that sounded distinctly inhuman. Natural, spontaneous human conversation is rarely, if ever, perfectly fluent.³⁹ It is peppered with breaks, irregularities, and non-lexical vocables collectively known as speech disfluencies.⁴⁰

These disfluencies include a variety of phenomena:

• Fillers: Sounds and words like "um," "uh," "ah," "er," "like," and "you know"⁴⁰
• False Starts: Sentences that are cut off mid-utterance and restarted
• Repetitions: The repeating of syllables, words, or entire phrases
• Repaired Utterances: Self-corrections, such as "He was wearing a black—uh, I mean a blue, a blue shirt"⁴⁰

In everyday, unscripted conversation, these elements are not rare occurrences; studies suggest they can appear at a rate of around 6 per 100 words, making up a significant portion of our spoken output.³⁹

The critical paradigm shift in modern TTS is the recognition that these so-called "imperfections" are not errors to be eradicated, but rather essential features to be modeled. They are fundamental to creating speech that sounds authentically human and conversational, distinguishing it from a voice that is merely reading a script. This leads to a fascinating paradox: the most "perfect" conversational AI voice is one that is perfectly imperfect. This embrace of organic, messy, and thoughtful speech represents the final frontier of realism, moving beyond mere accuracy to capture the true nature of human communication.

The Conversational Function of Fillers and Pauses

Far from being meaningless noise, disfluencies serve critical cognitive and interactive functions in dialogue:

• Signaling Thought and Planning: Fillers like "um," "hmm," or a prolonged "uh" are powerful signals to a listener. They indicate that the speaker is not disengaged but is actively thinking, planning their next words, or retrieving a piece of information from memory.⁴⁰ In a conversation, a long, silent pause might be misinterpreted as the end of a speaker's turn, inviting interruption. A filled pause, however, effectively holds the conversational floor, signaling "I'm not done yet, I'm just thinking".³⁹

• Improving Listener Comprehension: Research has shown that the presence of natural hesitations can actually improve a listener's comprehension and subsequent recall of the information being presented. This cognitive benefit goes beyond simply slowing down the speech rate, suggesting that these pauses prime the listener's brain for new or complex information.⁴²

• Conveying Politeness and Social Nuance: Fillers can also act as powerful social lubricants. Using a hesitation can soften a direct refusal, making it sound more considered and polite. For instance, responding to an invitation with "Hmm, sorry, no" is perceived as much more thoughtful than an abrupt "no".⁴⁰ They can also be used to hedge a statement, signaling uncertainty or a desire not to sound overly authoritative.⁴²

Programming "Ums" and "Ahs": The Final Frontier

Successfully modeling these natural disfluencies is a complex undertaking. It requires a fundamental shift in training data, moving away from clean, professionally read audiobooks and toward large, diverse corpora of spontaneous, conversational speech where these phenomena naturally occur.⁴³

Modern TTS platforms are approaching this challenge in two primary ways:

• Explicit Prompting: Many of the latest TTS models are capable of generating filler words if they are explicitly included in the input text. Developers using platforms like Deepgram Aura or ElevenLabs can simply write "um" or "uh" into their script, and the AI will speak the word, giving them direct, granular control over the placement of these disfluencies.⁴⁵ This approach is effective for crafting specific conversational flows.

• Implicit Generation: The most advanced and ambitious approach involves training models to generate these disfluencies automatically and contextually. Google's "Chirp 3: HD voices" are a prime example, explicitly designed to incorporate "human disfluencies" to create more natural and spontaneous conversational speech.¹⁰ This is achieved by using stochastic, attention-based neural models that are not deterministic. Instead of producing the exact same output every time, they have a degree of randomness that allows them to learn the statistical patterns of when and where a human speaker would likely pause or use a filler, and then replicate that behavior in their output.⁴³

VocalCopycat's Edge: Our philosophy of pursuing authentic, natural-sounding speech extends to these crucial conversational markers. Our proprietary neural models are trained on diverse datasets that encompass the natural rhythms and hesitations of real-world speech. This allows VocalCopycat to produce voices that not only speak fluently but also pause, hesitate, and "think" like a real person. This provides a level of conversational realism that is a step above the competition, all while rigorously upholding our core promise of delivering pristine, artifact-free audio.

Section 4: The Creator's Dilemma: Navigating the Landscape of Digital Voices

The Promise and the Pitfall: Quality vs. Artifacts

The modern landscape of Text-to-Speech technology is populated by a host of powerful and sophisticated tools. Major cloud providers like Google, Amazon, and Microsoft, alongside innovative and specialized startups such as ElevenLabs, offer an impressive array of capabilities.⁴⁷ Developers and content creators now have access to vast voice libraries spanning hundreds of languages, real-time audio streaming, deep customization via SSML, and even the ability to create unique custom voices.¹⁰

However, despite this technological progress, a persistent and deeply frustrating challenge remains for those who rely on these tools for professional content: the artifact problem. While many platforms promise high-quality, human-like output, they often produce subtle—and sometimes glaring—digital imperfections. These artifacts can manifest as a metallic sheen, an unnatural warble, sudden shifts in tone, or other strange glitches that shatter the illusion of a human speaker. For podcasters, audiobook narrators, YouTubers, and e-learning developers, this means hours of tedious post-production work, meticulously listening for errors, re-generating clips, and stitching audio together to create a seamless final product. This process turns creative professionals into audio engineers, disrupting their workflow and stifling their creativity.³

A Superior Solution for Flawless Audio: Introducing VocalCopycat

We built VocalCopycat with a singular mission: to solve this specific problem and democratize access to truly professional, broadcast-ready voice synthesis that works perfectly, right out of the box.³ We identified the gap between the promise of AI voice and the reality of the creator's workflow and engineered a solution from the ground up to bridge it.

Our platform is built on three core principles that directly address the pain points of modern content creators:

• Superior Quality, Fewer Artifacts: Our proprietary neural voice technology is the cornerstone of our platform. It has been meticulously engineered to dramatically reduce the digital artifacts that plague other TTS systems. By developing more sophisticated models that have a deeper understanding of natural human speech patterns, VocalCopycat delivers consistently clean, high-fidelity audio. Our users report up to a 70% reduction in the time spent on post-production editing compared to other leading voice AI platforms, allowing them to move from concept to published content faster than ever before.³

• More Efficient Voice Cloning: Creating a unique, branded voice is a powerful way to stand out. However, the process is often resource-intensive, with some competitors requiring 30 minutes or more of pristine, studio-quality audio to create a convincing clone.³ VocalCopycat's advanced models are significantly more data-efficient, enabling the generation of high-quality, convincing voice replicas from much shorter audio samples. This lowers the technical and logistical barriers to entry, making custom voice creation a practical and accessible option for a much wider range of creators.³

• Creator-First Pricing: We believe that premium technology should not be locked behind an enterprise-level paywall. The prevailing subscription-based, character-limited pricing models of many competitors can become prohibitively expensive for creators with high-volume needs.⁴⁵ VocalCopycat offers a transparent, affordable, and straightforward pricing structure with one-time payment packages. This approach provides exceptional value and predictability, ensuring that top-tier voice synthesis is accessible to independent creators, small businesses, and large enterprises alike.³

Comparative Analysis: Choosing the Right Tool for the Job

To provide a clear, at-a-glance summary of the market, the following table compares the leading TTS platforms across the critical elements of quality, expressiveness, and efficiency discussed throughout this report. This analysis is designed to help creators and developers make an informed decision based on the factors that matter most for producing high-quality audio content.

<google:style>

Conclusion: The Future is an Artifact-Free, Expressive Voice

We have journeyed through the intricate layers that constitute a truly realistic AI voice, deconstructing it from its foundational rhythm of prosody and intonation, to the rich color of emotional inflection, and finally to the authentic, humanizing stumbles of disfluency. It is clear that achieving genuine realism is a monumental task of both science and art, requiring sophisticated models trained on vast and diverse datasets.

The next evolution in Text-to-Speech, however, is not simply about adding more features or voices. It is about perfecting the final output. The new standard of quality against which all platforms will be measured is the ability to deliver audio that is not only expressive, nuanced, and natural, but also impeccably clean, consistent, and free of the digital artifacts that shatter immersion and create hours of remedial work for creators.³

This is the promise of VocalCopycat. We believe that advanced voice technology should enhance creativity, not hinder it. By engineering a solution that tackles the artifact problem head-on and by making high-fidelity voice cloning more efficient and affordable, we empower you to focus on the art of creation. We handle the complex science of speech synthesis so that you can produce compelling content that captivates your audience without compromise. Your creative flow can finally be unleashed.

Experience the VocalCopycat difference for yourself. Your voice matters, and with our platform, artifacts don't. Sign up today to claim your free 2,000 credits and take advantage of our limited-time introductory pricing. Start creating flawless, broadcast-ready audio in minutes.⁵⁵

Works Cited

1. Text to Speech with Emotion: A Game Changer for Creators in 2025 - CAMB.AI, accessed June 24, 2025, https://www.camb.ai/blog-post/text-to-speech-with-emotion
2. AI Voice Generator and Text-to-Speech Tool - Amazon Polly, accessed June 24, 2025, https://aws.amazon.com/polly/
3. VOCALCopyCat: Voice generation and cloning cheaper and better than 11Labs | Product Hunt, accessed June 24, 2025, https://www.producthunt.com/products/vocalcopycat
4. Vocalcopycat is now on Microlaunch, accessed June 24, 2025, https://microlaunch.net/p/vocalcopycat
5. Macquarie University, accessed June 24, 2025, https://www.mq.edu.au/faculty-of-medicine-health-and-human-sciences/departments-and-schools/department-of-linguistics/our-research/phonetics-and-phonology/speech/phonetics-and-phonology/intonation-prosody
6. Intonation - prosody - Macquarie University, accessed June 24, 2025, https://www.mq.edu.au/faculty-of-medicine-health-and-human-sciences/departments-and-schools/department-of-linguistics/our-research/phonetics-and-phonology/speech/phonetics-and-phonology/intonation-prosody
7. Speech Prosody: The Musical, Magical Quality of Speech - Frontiers for Young Minds, accessed June 24, 2025, https://kids.frontiersin.org/articles/10.3389/frym.2021.698575
8. Synthesizing Intonation - Phonetics Laboratory, accessed June 24, 2025, https://www.phon.ox.ac.uk/jpierrehumbert/publications/synthesizing_intonation.pdf
9. Speech SDK How to get proper intonation when synthesizing a partial sentance., accessed June 24, 2025, https://learn.microsoft.com/en-au/answers/questions/1538167/speech-sdk-how-to-get-proper-intonation-when-synth
10. Text-to-Speech AI: Lifelike Speech Synthesis - Google Cloud, accessed June 24, 2025, https://cloud.google.com/text-to-speech
11. How is prosody controlled in modern TTS systems? - Zilliz Vector Database, accessed June 24, 2025, https://zilliz.com/ai-faq/how-is-prosody-controlled-in-modern-tts-systems
12. Voice and sound with Speech Synthesis Markup Language (SSML) - Learn Microsoft, accessed June 24, 2025, https://learn.microsoft.com/en-us/azure/ai-services/speech-service/speech-synthesis-markup-voice
13. How is prosody controlled in modern TTS systems? - Milvus, accessed June 24, 2025, https://milvus.io/ai-quick-reference/how-is-prosody-controlled-in-modern-tts-systems
14. Milvus.io, accessed June 24, 2025, https://milvus.io/ai-quick-reference/how-is-prosody-controlled-in-modern-tts-systems
15. How is prosody generated in TTS outputs? - Milvus, accessed June 24, 2025, https://milvus.io/ai-quick-reference/how-is-prosody-generated-in-tts-outputs
16. Speech synthesis: The path to creating expressive text-to-speech, accessed June 24, 2025, https://inworld.ai/blog/speech-synthesis
17. Emotional Speech Synthesis - USC Institute for Creative Technologies, accessed June 24, 2025, https://people.ict.usc.edu