Artificial intelligence
Text to speech IA
Publiée le October 19, 2025
Publiée le October 19, 2025
Artificial intelligence applied to text-to-speech (TTS) refers to all models and services capable of transforming text into natural speech. Long seen as an assistive technology for the visually impaired, text-to-speech has gained strategic importance with the rise of conversational agents and virtual assistants. Advances made between 2010 and 2020, in particular the adoption of deep neural networks, have seen TTS evolve: the voice generated is no longer robotic, but expressive, multi-lingual and available in real time.
This study offers a complete overview of TTS in 2025: definition, model architecture, presentation of the main commercial and open source models, evaluation criteria, use cases, ethical issues and future trends. The aim is to offer a richer, more structured article than the first research results available.
Speech synthesis refers to a technology that converts text into an intelligible, natural audio signal, often referred to as speech synthesis or reading aloud. IBM points out that TTS is a tool that “transforms text on a digital interface into natural audio”, and was developed as an assistive technology. As early as the 1930s, experimental electric synthesizers appeared; in the late 1950s, algorithms based on recorded sound bases assembled syllables, but the voices generated remained monotonous. The arrival of neural networks in the 2000s marked a turning point: models learn waveforms directly, producing realistic voices. Today’s voice AI generators use thousands of hours of recordings to train expressive, multilingual voices, surpassing parametric and concatenative systems.
An important milestone is WaveNet, a model proposed in 2016 by DeepMind. WaveNet is a fully probabilistic, autoregressive generative neural network that predicts each audio sample from previous ones. This architecture generates voices considered more natural than conventional parametric synthesizers; a single model can mimic several voices based on the speaker’s identity.
Since 2023, the most striking development has been the multiplication of open models (XTTS, Kokoro, Orpheus…) capable of competing with commercial APIs. Platforms such as Layercode show that the quality of open source models is increasing so fast that the difference between them and the market leaders is disappearing.
Voice generation comprises two main phases: linguistic analysis of the text, followed by synthesis. After receiving a text, the system breaks down sentences, identifies abbreviations and converts numbers into words. Prosodic analysis estimates rhythm, intonation and pauses; it determines pronunciation according to context and prepares conversion into phonemes. IBM points out that the neural networks receive audio corpora and their transcriptions in order to understand the link between words, accents, tonality and rhythm.
Once the linguistic analysis has been completed, the synthesis is generally carried out in two stages. First, the model converts the phoneme sequences into temporal representations (such as spectrograms) that describe the variation of frequencies over time. Next, a neural vocoder reconstructs the sound wave from the spectrogram. This phase is crucial: vocoders such as WaveNet, HiFi-GAN or WaveGlow transform the spectral representation into natural sound by directly modeling the audio wave.
In cloud APIs, synthesis can take place in streaming mode (conversion as the text arrives) or in delayed mode for long passages (e.g. audiobook reading). Developers often measure Time-To-First-Byte (TTFB) – the time it takes to receive the first audio block. Layercode notes that, for natural interactions, TTFB should remain under 200 milliseconds. Some real-time models such as Cartesia Sonic or ElevenLabs Flash prioritize latency over prosody, while high-fidelity models (Dia 1.6B, Coqui XTTS) analyze full text to optimize intonation and emotion.
The detailed operation of a commercial API follows the pipeline described by Vonage:
Input and pre-processing: the API receives a text, normalizes dates, numbers and abbreviations, then segments the sentences.
Linguistic analysis: the system establishes the syntactic structure and adds prosodic information (intonation, accentuation) according to the context.
Phonetic conversion: text is translated into a sequence of phonemes, the basic unit of sound.
Prosody generation: a model generates the rhythm, pitch and duration of sounds to reflect the desired emotion.
Speech synthesis: a vocoder (concatenative, parametric or neural) constructs the audio waveform. Modern solutions are based on deep neural networks.
Audio playback: the API returns an audio stream or sound file. APIs cache frequent phrases to reduce latency.
The use of XML Speech Synthesis Markup Language (SSML) enables developers to control speed, tone and volume, or to combine several voices in the same text.
The 2025 market is divided between commercial solutions and open source models. The multiplication of open-source models makes competition fierce. The following table summarizes some of the major solutions (see image below). [Insert image of text-to-speech models table here].
IBM Watson Text to Speech: cloud service offering standard or neural voices, with over 20 languages and the option of creating a personalized voice. The API supports streaming via WebSocket or REST, and provides MP3 or WAV formats.
Microsoft Azure Speech TTS: a platform for realistic neural voices with advanced features. It enables real-time synthesis via SDK or REST, asynchronous synthesis for long texts, personalized voices, and the use of SSML to adjust prosody. Azure supports visemes to synchronize speech with facial animation.
Google Cloud Text-to-Speech: service offering standard, WaveNet and Neural2 voices in over 40 languages. WaveNet models produce a natural voice by predicting the audio wave sample by sample, and are used in many Google Assistants. Neural2 voices, announced in 2025, improve prosody and support additional languages. The platform offers SSML control and per-character pricing.
Amazon Polly: AWS API offering standard and neural voices in over 30 languages. Polly stands out for its vocabulary customization and the ability to adjust pronunciation via phonetic dictionaries. It also offers caching functionality to reduce latency.
Deepgram Aura-2: targeted at call centers. Aura-2 guarantees a TTFB of less than 200 ms and billing to the letter, but offers only two languages and does not support voice cloning.
Coqui TTS: modular toolbox implementing several architectures (Tacotron 2, FastSpeech, Glow-TTS, VITS). It offers multi-speaker and multilingual models with over 1,100 languages, and lets you train or customize your own voice. The project is licensed under the Mozilla Public License (MPL).
Coqui XTTS v2.0.3: high-fidelity model capable of producing moving voices in several languages. Ideal for narration, it processes the entire text to optimize prosody. It supports voice cloning from a few seconds of recording.
Canopy Labs Orpheus: a family of open source models (3 Md, 1 Md and 400 M parameters) offering a compromise between quality and performance. Orpheus offers multilingual voice cloning and latency adapted to streaming. According to Layercode, Orpheus rivals the commercial leaders in terms of naturalness.
Hexgrad Kokoro: 82 M-parameter real-time model that prioritizes speed. It is designed for conversational agents where latency must be kept to a minimum.
Dia 1.6B (Nari Labs): high-fidelity model with 1.6 billion parameters. It offers expressive voices and multilingual support, but generation is slower than with real-time models.
Chatterbox: a small open-source model based on the Llama 0.5 B family. According to Modal and Layercode, it is optimized for speed and simplicity, and provides a gateway for novice developers.
Tacotron 2: sequence-to-sequence architecture introduced by Google, combining a text encoder and an attentive decoder that produces a spectrogram. It has served as the basis for many open source models. Tacotron 2 improves prosody over Tacotron 1, but requires a vocoder like WaveNet to convert the spectrogram into audio.
FastSpeech: parallel flow model that generates spectrograms faster by predicting phoneme duration. FastSpeech speeds up synthesis and is still widely used in real-time applications.
VITS (Variational Inference with adversarial learning for TTS): all-in-one architecture that integrates spectrogram generation and vocoding in a single end-to-end trained model. VITS produces a natural voice and offers competitive synthesis times.
Glow-TTS: an invertible flow model based on flow transformations that allows explicit control of prosody. It offers a compromise between quality and speed.
According to Modal, the evaluation of speech synthesis must take into account five axes:
Naturalness: quality perceived by the listener. Comparison platforms such as TTS Arena use human votes to assess the naturalness of models.
Voice cloning capability: the ability to reproduce a voice from a few seconds of recording. Cloning is essential for creating branded voices or customized characters.
Word Error Rate (WER ): measures the accuracy of reverse transcription by a speech recognition system. A low WER means that the synthesized speech is well understood.
Latency: response time, measured by the TTFB for real-time applications or by the RTFx factor for offline synthesis. High latency can make interactions unnatural.
Number of parameters: size of the model, which influences the resources required and the cost. A large model (e.g. 5.77 billion parameters for Higgs Audio V2) requires a substantial GPU infrastructure.
Fingoweb also recommends examining voice quality, language support, personalization, speed and integration with other tools.
The use of TTS has spread to many fields:
Accessibility and education: reading aloud for the visually impaired, dyslexics or foreign language learners. E-learning platforms use TTS to improve engagement and memorization.
Voice assistants and chatbots: Siri, Alexa and Cortana rely on speech-to-text / text-to-speech loops to converse with users. TTS broadcasts messages, notifications and options in voice interfaces.
Customer service and call centers: modern IVRs leverage TTS APIs to direct calls, present menus and answer questions. The low latency of real-time models is essential to avoid silences.
Audio and marketing content: narration of articles, automatically generated podcasts, presentation videos, voice-overs for e-learning or advertising. TTS enables creators to rapidly transform text into audio in multiple languages.
Health: medication reminders and reading of medical records for patients and caregivers.
Documentation and compliance: audio document generation for meetings, audio transcriptions for training and archiving.
Speech synthesis raises questions of ethics and responsibility:
Deepfakes: the ability to clone voices from a few seconds of audio can be misused to imitate a person without their consent. IBM notes that the rise of TTS has led to controversy surrounding deepfakes, and that detection techniques are currently being developed. Suppliers need to put systems in place to authenticate voices and prevent abuse.
Privacy: training a model on human voices requires sensitive data. Companies must obtain informed consent and anonymize recordings to comply with regulations.
Language bias and accents: some models favor English or dominant accents, which disadvantages minority languages or dialects. The rise of multilingual models aims to reduce these biases, but quality varies from language to language.
Energy cost: large models consume a lot of energy for training and inference. Model selection must balance performance and environmental footprint.
Real-time and ultra-low latency: the boundary between human conversation and text-to-speech is shrinking. Models such as ElevenLabs Flash v2.5 offer TTFB of less than 100 ms for 30 languages. Future versions aim to go below 50 ms.
Personalization and expressive cloning: the integration of high-fidelity voice cloning into consumer platforms (ElevenLabs, Coqui XTTS) democratizes the creation of branded voices or fictional characters. The models support emotional intonation and multilingual generation from a single voice.
Multimodal integration: new models, such as GPT-4o mini, combine text, images and audio. They can control prosody via prompts and synchronize speech with animation (visemes).
Mature open source: the open source ecosystem has reached a level of maturity that allows it to be deployed in production. Models such as XTTS-v2.0.3, Orpheus or Dia rival commercial APIs in terms of quality and cost, while modular frameworks simplify customization. Developers prefer openness to avoid vendor dependency.
Regulation and fraud detection: the massive adoption of TTS is prompting governments to set standards for voice authenticity and sanction deepfakes. New detection techniques based on acoustic fingerprinting or digital signatures are currently being deployed.
Artificial intelligence applied to speech synthesis is booming: today’s computer-generated voice is fluid, expressive and virtually indistinguishable from that of a human. This progress is the result of the integration of deep neural networks, innovative vocoders and gigantic audio corpora. Commercial APIs (IBM, Microsoft, Google, Amazon…) and open source models (Coqui, Orpheus, XTTS) offer a range of solutions to suit every need, from ultra-low latency for conversational agents to studio quality for podcasts. However, these advances are accompanied by ethical issues linked to voice cloning and data protection.
When choosing a model or service, it’s worth looking at naturalness, cloning capability, error rate, latency and model size. Trends in 2025 show an emphasis on personalization, real-time integration and open code. Speech synthesis, once an accessibility tool, is becoming an essential component of digital communication and immersive user experiences.