# Text-to-Speech for Phone Calls: Voice, Speed and Pronunciation

> How to make a phone AI easy to understand: picking a voice for narrowband audio, setting speaking rate, writing text TTS reads correctly, and testing on calls.

Published: 2026-11-10
Source: https://voicast.jp/en/blog/text-to-speech-for-phone

## Key points

- Phone audio is narrowband, so a voice that sounds great on your laptop can turn muddy on a call. Pick voices by listening over an actual phone line.
- Too slow feels laggy, too fast causes repeats. Run slightly brisk overall and slow down by chunking whenever you read out numbers, dates or times.
- Fix pronunciation in the text you write, not in the engine. Spell out times and dates, avoid symbols and abbreviations, and respell ambiguous words.

Getting text-to-speech right for phone calls comes down to three choices: the voice, the speaking rate and the way you write the text it reads. Make all three by listening over a real phone line, because phone audio strips out much of what makes a voice sound good on your laptop. And most mispronunciations are cheaper to fix by rewriting the text than by tuning the engine.

## What phone audio does to a voice

A phone call carries far less audio than a podcast or a video. Keep that in mind before you audition voices:

- **It's narrowband.** Twilio Media Streams, for example, carries call audio as 8,000 Hz μ-law ([Twilio docs](https://www.twilio.com/docs/voice/media-streams/websocket-messages)). The high frequencies are gone, so sounds like "s", "f" and "th" blur together.
- **Listeners are in noisy places.** Cars, sidewalks, speakerphone in a busy kitchen.
- **Voice is the only channel.** No captions, no screen. If they miss something, all they can do is ask you to repeat it.

So for phone calls, choose for clarity first and polish second.

## Choosing a voice

Weigh intelligibility and brand fit together:

| Consideration | Works well | Avoid |
|---|---|---|
| Intelligibility | Clear, crisp, well-articulated voices | Breathy or raspy voices, which sink into line noise |
| Brand fit | Calm for healthcare and legal; brighter for restaurants and retail | Anything that clashes with the reason people call |
| Long explanations | Even, explanatory delivery | Very animated delivery, which gets tiring quickly |

Shortlist two or three voices and have each read the same script over the phone. Ask the people who answer the phones today to listen too. They know your callers better than the team that built the bot.

## Setting the speaking rate

Both extremes cost you on the phone:

- **Too slow** feels like waiting. Combined with normal response latency, it makes the whole conversation drag.
- **Too fast** triggers "sorry, can you say that again?", especially from older callers or noisy environments, and repeats make calls longer, not shorter.

A practical pattern: keep the overall rate slightly brisk, and slow down the parts that matter by breaking them into short chunks. "Your appointment is Friday December 19th at 3 p.m." is harder to catch than "Your appointment is on Friday, December 19th. At 3 p.m."

## Writing text that reads well aloud

Most TTS mistakes trace back to text written for the eye. Write your fixed lines (greeting, closing, FAQ answers) to be heard.

| Written | Problem | Write instead |
|---|---|---|
| `9:00-18:00` | Symbols and 24-hour time read awkwardly | 9 a.m. to 6 p.m. |
| `12/19` | Date or fraction? US or UK order? | December 19th |
| `123 Main St.` | "St." may become "Saint" | 123 Main Street |
| `Dr. Lee, 4 Oak Dr.` | "Dr." is doctor or drive | Doctor Lee, 4 Oak Drive |
| `Suite 4B, 2nd fl.` | Abbreviations read letter by letter or not at all | Suite 4 B, second floor |
| `We'll read your file` | Heteronyms (read/read, lead/lead) can flip | Rephrase if it comes out wrong |
| `*`, `/`, `(…)`, `→` | Skipped or read literally, with odd pauses | Say it in words |

Keep sentences short. On the phone, people can't glance back at the start of a long sentence. Put one important fact (a date, a price, what to bring) in each sentence instead of stacking them.

Phone numbers deserve special care: group the digits the way people say them locally (3-3-4 in North America) and pause between groups. Japanese numbers are 10 or 11 digits starting with 0 and are grouped differently, so if you serve both, test both. More on reading values back in [read-back confirmation](https://voicast.jp/en/blog/confirm-by-reading-back).

## Testing on real calls

Don't sign off on a voice from a preview player. Test it on calls:

1. Set up a test number and run your agent in a test environment.
2. Call from a mobile phone in a quiet room.
3. Call again on speakerphone, and once more from outside.
4. Have the listener write down every date, phone number and price the AI reads out.
5. Fix the text wherever their notes don't match, change one setting (voice or rate), and repeat.

Changing one thing per round tells you what actually helped.

## Doing this with voicast

In voicast, the agent settings has two settings for this: `voice` picks one of 30 voices by ID, each named for its character, such as `female-firm` or `male-calm`. `speaking_rate` picks from fixed steps between `0.85` (very slow) and `1.5` (fastest), defaulting to `1.0` (normal).

```json
{
  "voice": "male-calm",
  "speaking_rate": 1.1,
  "greeting": "Thanks for calling Oak Street Dental. How can I help?"
}
```

You can preview each voice in the dashboard. Setups are versioned, so changing the voice or rate never affects calls already in progress; they keep the version they started with. The `domain` and `vocabulary` hints help speech recognition, not pronunciation. To fix how something is said, rewrite your greeting, closing or FAQ text.

## FAQ

### Is a male or female voice better for phone calls?

Choose on clarity, not gender. Have each candidate read the same script over the phone and keep the one listeners transcribe most accurately.

### How do I fix a name the voice keeps mispronouncing?

In fixed text such as the greeting or FAQ answers, respell it phonetically ("Nguyen" as "Win", for example) and test until it sounds right. Start with your business name and street names, since callers hear those most.

### Can the speaking rate change per caller?

In voicast the rate is set per agent settings and stays the same for the whole call. If one line serves mostly older callers, give that line its own setup with a slower rate.
