In a blind listening test of 4,000 people, Microsoft says 50.3% rated its new AI voices as equally or more human-like than real human recordings. Put simply, half the room could no longer tell the machine from the person.
That result came with the launch of MAI-Voice-2.1 on 1 October 2026, one of several new voice models released in the first week of the month. The technology behind it has a plain name: text-to-speech, often shortened to TTS.
By the end of this guide, you will know what text-to-speech is, how it went from robotic to convincing, what it costs, where a Hong Kong small business can use it today, and the mistakes worth avoiding.
What is AI text-to-speech?
AI text-to-speech is software that turns written words into spoken audio. You type or paste text, choose a voice and language, and the system produces a natural-sounding recording in seconds. Modern AI versions learn speech patterns from thousands of hours of human recordings, so they handle tone, pauses and emphasis rather than reading word by word.
The idea is not new. Your phone's sat-nav and the voice reading train announcements have used text-to-speech for years.
What changed is quality. Older systems stitched together tiny pre-recorded sound fragments, which is why they sounded flat and choppy. Today's AI systems generate the sound wave itself, much like an image generator paints a picture pixel by pixel.
Think of the difference between a ransom note made of cut-out magazine letters and a handwritten letter. Both carry the message. Only one feels like a person wrote it.
How does AI text-to-speech work?
AI text-to-speech works in three stages. First, it reads the text and works out how each word should be pronounced. Second, it plans the rhythm: where to pause, which words to stress, and whether a sentence is a question. Third, a neural network generates the actual audio wave, sample by sample, in the chosen voice.
Stage 1: Reading the text
The system expands things a human would read naturally. "HK$1,200" becomes "one thousand two hundred Hong Kong dollars". "Dr." becomes "doctor". This step is where many mistakes start, especially with names, addresses and mixed languages.
Stage 2: Planning the delivery
Next, the model predicts prosody, the music of speech. A sentence ending in a question mark rises. A list gets small pauses between items. A warning sounds firmer than a welcome.
Stage 3: Generating the sound
Finally, a neural network produces the waveform. Because it learned from real recordings, it adds the small breaths and natural timing that make a voice feel alive.
Speed has improved sharply. Microsoft states that its MAI-Voice-2.1-Flash model can generate 45 seconds of audio with an end-to-end delay of about 150 milliseconds, roughly the time it takes to blink. That speed is what makes live phone conversations with an AI possible.
What is the difference between text-to-speech, speech-to-text and voice cloning?
Text-to-speech turns writing into sound. Speech-to-text does the opposite, turning spoken words into written text, as meeting transcription tools do. Voice cloning is a special kind of text-to-speech that copies one specific person's voice from a short sample, which brings extra legal and security risks.
These three terms are often mixed up, so here is a simple way to remember them:
--- Text-to-speech (TTS): the AI's mouth. Words in, voice out.
--- Speech-to-text (STT), also called transcription: the AI's ears. Voice in, words out. Microsoft's MAI-Transcribe-2-Streaming, released the same day, covers 60 languages at an introductory price of US$0.54 per hour of audio until the end of 2026.
--- Voice cloning: a mouth that imitates a particular person. Microsoft says both new MAI-Voice models can clone a voice from a few seconds of reference audio, with built-in consent safeguards.
When an AI phone agent answers your hotline, it uses all of them together: ears to hear the caller, a language model to decide what to say, and a mouth to say it. If you want the security side of cloning, read our guide on what voice cloning is and the fraud risks it creates.
How much does AI text-to-speech cost?
Business text-to-speech is usually priced per million characters of text. Microsoft lists MAI-Voice-2.1 at US$22 per million characters and the faster Flash version at US$15. One million characters is roughly 18 hours of English narration, so a typical small business spends a few dollars a month, not hundreds.
Here is a rough way to estimate it. A calm speaker says about 150 words a minute, and an English word averages around six characters including the space. That is about 900 characters a minute, or roughly 1,100 minutes of audio per million characters.
Some real-world examples at the US$22 rate:
--- A 60-second phone greeting (about 900 characters): far less than one US cent to generate.
--- Ten 2-minute product videos a month (about 18,000 characters): roughly US$0.40.
--- An AI hotline speaking for 50 hours a month (about 2.7 million characters): roughly US$60, before the cost of the listening and thinking parts.
Consumer apps price differently, usually as a monthly subscription with a character or minute allowance. Always check whether commercial use is allowed on the plan you choose, because many free tiers forbid it.
How can a Hong Kong small business use text-to-speech?
Hong Kong SMEs use text-to-speech for phone greetings and hold messages, voiceovers for social media videos, multilingual announcements in shops and clinics, audio versions of training material, and as the voice of AI phone or WhatsApp voice agents. The main benefit is changing a recording in minutes without booking a studio or a voice artist.
A trading company's phone menu
Office hours change for a typhoon or a public holiday. Instead of re-recording, the manager edits one line of text and generates a new Cantonese and English greeting in under a minute.
A café's short videos
The owner writes a 30-second script for a new seasonal drink and generates a voiceover in Cantonese, English and Mandarin, so one video serves locals, expats and mainland visitors.
A clinic's reminders
Front-desk staff spend hours calling patients. A recorded reminder in a clear, friendly voice can handle routine confirmations, leaving staff free for patients with real questions.
Training a new hire
A retail shop turns its written opening checklist into a 5-minute audio guide staff can listen to on the MTR.
Language matters here. Microsoft says MAI-Voice-2.1 covers 23 languages and 26 locales, and can keep one voice across languages with a native accent in each. For Hong Kong, always test Cantonese specifically. Cantonese has six main tones, and a wrong tone can change the meaning of a word entirely. Major cloud providers, including Amazon and Microsoft, do offer Hong Kong Cantonese voices, but quality varies by voice.
What are the common misconceptions about AI voices?
The biggest misconceptions are that AI voices still sound robotic, that customers will always notice, that any voice online is free to use commercially, and that a good voice means the content is correct. Text-to-speech only reads what it is given. It cannot check facts, prices or promises in your script.
Myth 1: "AI voices still sound like robots."
That was true five years ago. Microsoft's 4,000-listener test is a vendor claim, but independent listening across the industry points the same way: the best voices are now hard to spot in short clips.
Myth 2: "A natural voice means the message is right."
A confident voice reading a wrong price is worse than a robotic one, because people believe it. Every script needs a human check before it goes live.
Myth 3: "I can copy any voice I like."
Imitating a real person, such as a celebrity or a former employee, without permission can raise legal, privacy and reputational problems. Use the stock voices a provider licenses to you, or a voice from someone who has agreed in writing.
Myth 4: "Customers do not need to know."
Many businesses now tell callers when they are speaking to an automated system. Being upfront builds trust, and it protects you if a caller later disputes what they were told.
How do you choose a text-to-speech tool?
Choose a text-to-speech tool by testing your own scripts in Cantonese, English and Mandarin, checking the commercial licence, comparing price per million characters or per minute, confirming how fast it responds for live calls, and asking where your text is processed and stored. A 20-minute side-by-side test usually reveals the winner.
Use this short checklist:
--- Test real content. Use a script with your shop's name, a street address, a price and a phone number. These are where voices stumble.
--- Check the tones. Ask a native Cantonese-speaking colleague to listen for wrong tones and awkward pauses.
--- Read the licence. Confirm commercial use is allowed for ads, phone lines and social media.
--- Match speed to the job. A video voiceover can take a few seconds. A live phone agent needs replies in well under a second.
--- Ask about data. If scripts contain customer names or medical details, check where the text is processed, in line with Hong Kong's Personal Data (Privacy) Ordinance.
--- Start with one use. A phone greeting or a single video series is enough to learn what works.
Frequently asked questions about text-to-speech
Is AI text-to-speech free?
Many tools offer a free tier with a small monthly allowance, but free plans often ban commercial use. Paid business pricing is usually per million characters, such as Microsoft's US$15 to US$22, which works out to a few dollars a month for most small businesses.
Can AI text-to-speech speak Cantonese?
Yes. Several major providers offer Hong Kong Cantonese voices. Quality varies more than in English, so test tones, names and addresses with a native speaker before using any voice with customers.
Is it legal to use an AI voice in my ads?
Using a provider's licensed stock voice on a plan that permits commercial use is generally fine. Copying a real person's voice without written consent is where the legal and reputational risks begin.
Will text-to-speech replace my staff?
On its own, no. Text-to-speech only reads text aloud. It becomes part of a customer-service system only when combined with speech recognition and an AI that decides what to say, and even then it works best on routine questions.
What should a business owner take away?
AI text-to-speech has crossed the line from robotic to convincing, and it now costs cents rather than studio fees. The smart first step is one small, low-risk use, such as a phone greeting, tested carefully in Cantonese and checked by a human before customers ever hear it.
The voice is the easy part now. The harder part is deciding what your business should say, when, and to whom. Get the script right, and the technology will do the reading.
That is the spirit we bring to every project: We understand AI. UD stands with you.
Reviewed by the UD AI team.
Curious where AI voices, phone greetings or a full AI receptionist could fit in your business? Take the free AI Ready Check and we will walk you through it step by step, from picking the first low-risk task to testing it in Cantonese before any customer hears it.
