AI voice and speech

Traditional text-to-speech systems worked by stitching together pre-recorded fragments of human speech (individual sounds, syllables, or words) and playing them back in sequence. This is why they sounded choppy and unnatural. The transitions between fragments were always slightly off, and the system had no understanding of meaning or context, so emphasis and intonation were essentially random.

Modern AI text-to-speech works differently. Models like those used by ElevenLabs are trained on vast amounts of human speech data. They learn not just how individual words sound but how people actually talk: the rhythm of a sentence, where emphasis naturally falls, how tone shifts with punctuation, how a question sounds different from a statement. When you give these models a piece of text, they generate speech from scratch rather than stitching together fragments. The result is fluid, natural-sounding audio that captures many of the subtle qualities of human speech.

The quality gap between the best AI voices and obviously synthetic speech is now enormous. The best models can convey warmth, urgency, and hesitation in ways that are often hard to distinguish from a human recording, at least for short passages. Longer passages sometimes reveal patterns: a slight uniformity in pacing, a tendency to handle unusual words or abbreviations oddly, or an emotional flatness that creeps in when there's no obvious tone to match.

Speech-to-text (also called transcription or speech recognition) faces a different challenge: making sense of the messy, imperfect way humans actually talk. We mumble, we speak over each other, we change direction mid-sentence, we use filler words, and we rely heavily on context that isn't in the words themselves.

AI transcription models, the most influential being OpenAI's Whisper, are trained on hundreds of thousands of hours of transcribed audio. They learn to handle accents, background noise, specialist vocabulary, and the general chaos of spoken language. Modern systems achieve accuracy rates above 95% for clear speech in quiet environments, and they're getting better in noisier conditions too.

Where current transcription tools still struggle is with heavy accents, overlapping speakers, very technical terminology, and audio quality that's genuinely poor. They also make mistakes that a human transcriber wouldn't, occasionally substituting a similar-sounding word that makes no sense in context, or mishearing proper nouns. These errors are worth looking out for in your activity.

ElevenLabs (elevenlabs.io) is the standout tool in this category. It offers the most natural-sounding AI voices currently available to consumers, with a library of thousands of voices in over 70 languages. The free tier gives you 10,000 characters per month, roughly equivalent to ten minutes of generated audio. That's more than enough for this activity, though you'll want to be thoughtful about how you use your allowance rather than generating dozens of test clips. You can sign up with email, Google, or Apple. The free tier doesn't include voice cloning or commercial usage rights, but it gives you full access to the voice library and the core text-to-speech engine.

Natural Reader (naturalreaders.com) offers a straightforward free tier with a selection of AI voices. It's less sophisticated than ElevenLabs (the voices are good but noticeably more "AI" in character) but it's simple to use and doesn't require an account for basic generation. It's a solid backup if you want to compare voices across platforms.

Your phone's built-in text-to-speech is worth trying as a baseline comparison. Both iOS (Settings > Accessibility > Spoken Content) and Android (Settings > Accessibility > Text-to-Speech) have built-in speech engines that have improved significantly in recent years. Comparing these with a dedicated AI voice tool is a quick way to appreciate how far the technology has come.

ChatGPT's Read Aloud feature allows you to have any ChatGPT response spoken aloud using OpenAI's voice models. It's not a standalone text-to-speech tool, but the voice quality is high and it gives you another point of comparison if you're already using ChatGPT.

Otter.ai (otter.ai) is a popular transcription tool with a free tier that includes a limited number of transcription minutes per month. It's designed primarily for meeting transcription (it can identify different speakers, highlight key points, and generate summaries) but it works perfectly well for transcribing a simple voice recording. Available on web and mobile.

Your phone's built-in dictation is likely the most accessible option you already have. On iPhone, tap the microphone icon on any keyboard. On Android, use Google Voice Typing (the microphone on the Gboard keyboard). Both have become surprisingly accurate for general dictation, and they work offline on most recent devices. The advantage is zero setup; the limitation is that they're designed for dictation (speaking in a structured way) rather than transcribing natural, conversational speech.

Google Docs voice typing (Tools > Voice typing in any Google Doc) provides free speech-to-text directly in your browser. It requires a Chrome browser and an internet connection but handles extended dictation well.

Microsoft Word dictation offers similar functionality built into Word Online and the desktop app (Home > Dictate). If you already use Microsoft 365, this is worth trying.