Every speaking voice belongs to one language and one quality tier, so you pick a voice by its language code first and its tier second, and keep the agent’s language, tier and voice name in agreement. The language your AI employee listens in is a separate setting from the voice it speaks with, so a Malay voice does not by itself make it listen for Malay.
Think of a receptionist with two skills you hire for separately: how well they understand callers, and how they sound on the phone. A lovely voice tells you nothing about their ear.
How do I pick a voice for a language?
Start from the language code, then choose a tier, then a voice name from the same row of the speech provider’s published list. A voice name is built as <language>-<tier>-<name>, so it already carries both the language and the tier. Three agent settings have to agree with it:
ttsLang: the language, such asms-MYoren-USvoiceTier: the quality tierttsVoice: the full voice name
If they disagree, the request may be rejected or the voice ignored; do not rely on either. Wetel keeps no list of its own, because the provider adds and retires voices on its own schedule. The agentConfigOptions query returns links to both catalogues and a short curated voice list.
If you only want a natural default, pick a tier and a language and leave the voice name alone. A workspace can also set default voice settings once; a value set on the agent always takes precedence.
What do the tiers trade off?
There are five tiers, from cheapest and clearly synthetic to a narration tier that is not meant for live chat. As the docs describe them:
| Tier, in order | Use it for | Relative cost |
|---|---|---|
| The lowest tier | Cheapest, clearly synthetic | Lowest |
| The second tier | Better quality, still budget | Low |
| The middle tier | Natural; a safe default | Medium |
| A higher tier | The most natural for live conversation | About 2x the middle tier |
| Narration tier | Recorded content; not for real-time chat | About 10x the middle tier |
The exact tier values are in the guide’s text-to-speech section. “Natural” is the docs’ description, not a score we measured.
How is the listening language set?
With a primary language and up to three alternatives, all separate from the voice. The primary has by far the biggest effect on what the listener hears correctly; if you leave it empty, the agent’s speaking language is used. The alternatives are extra languages it may switch to:
- Maximum three. The underlying engine ignores the whole list above three, so Wetel rejects it up front.
- Defaults: when you set nothing, the alternatives are
en-USandms-MY. An explicit empty list means no alternatives, which is not the same as leaving it unset. - Chinese is not in the defaults on purpose. It caused accented English to come back as Chinese. If you expect Chinese, make it the primary or name it yourself.
- Not supported at any setting: Hokkien and Hakka.
There are two listening engines. One uses the primary and alternatives above. The other works out the language itself and ignores both settings. Which one your environment uses by default is in the speech service reference.
What about replies that mix languages?
Speaking mixed-language text is untested. A voice reads in its one language, so Malay words inside an English sentence may sound foreign. A slower route meant for recorded voice notes works out the language from the text and may do better, but we have not measured it. The docs’ advice is to have a native speaker listen to your own phrases on every voice you plan to ship.
For listening to mixed speech, see voice notes in English, Malay and Chinese.
What have we not measured?
We have not measured accuracy on recordings of your users. Our own tests used generated speech and openly licensed public recordings, not spontaneous Malaysian speech that mixes Malay, English and Chinese. So treat any setup as a starting point: collect real clips, write the correct text by hand, and compare before you go live.
Whatever you pick, the workspace owner decides which voice speaks for the business, and a person should listen before customers do.
Related reference docs
- Choosing Languages and Voices: primary and alternative languages, tiers and voice names
- Speech-to-Text and Text-to-Speech Integration Guide: the two listening engines, mixed-language speech and testing on your own audio
- Avatar Config: workspace-wide voice defaults and how agent settings override them
- Sending Voice Messages to an Agent: transcribing a voice note and passing it to an agent
To set up your first agent’s voice, start with Choosing Languages and Voices.