AI Automation · Illustrated edition
Should the audio guide speak before the visitor arrives?
Zyphra’s ZONOS3-Mini can synthesise speech in a browser. A museum audio guide shows when local generation is useful and when approved recordings work better.

The short answer
Generate audio in advance when the script and its variations are known. Consider on-device synthesis when text genuinely changes during a visit, you can prepare the hardware, and a person has approved the voice and pronunciation policy. A browser demo does not settle first-use download time or quality on your visitors’ phones.
A recording can be synthetic and still be approved before anyone hears it. The important choice is when you commit the words to sound.
In this article
01 / 04
The visitor came to hear the instrument, not download a model
In an illustrative museum gallery, a visitor scans the label beside a brass astrolabe. The guide offers a short introduction or a longer explanation of how the rotating rule works. If those are the only two choices, the museum can approve two recordings and serve the one the visitor taps. There is no good reason for every phone to manufacture the same sentence.
Now change the brief. A curator wants to assemble a different route every morning from approved descriptions and new loans. That makes speech generation worth investigating, but not necessarily in the gallery. Rendering the day’s audio before the doors open may give the team all the flexibility it needs.
02 / 04
A smaller voice engine changes where the work can happen
Zyphra announced ZONOS3-Mini on 9 November 2026; the authors submitted their technical report on 12 November. This is a documentary analysis of that release, not an Axiom Forge listening test. Compared with ZONOS2, the authors’ new codec represents each second with fewer time steps, and the model uses only part of its network for each step. There is less work to produce a passage. That matters because the supplier offers an actual browser runtime using WebGPU, alongside downloadable weights and native runtimes.
Speech synthesis reads supplied text. It does not translate a label, check the date engraved on an instrument or answer a visitor’s history question. If a supplier promises all three, ask to see the approved translation, the source of the answer and the speech output separately. A fluent reading can hide an incorrect script particularly well.
The authors describe a model covering 43 languages, with voice conditioning from a reference recording. Published weights carry an Apache 2.0 licence. Neither the coverage claim nor the model licence tells a museum whether a particular pronunciation is acceptable or a particular narrator agreed to this use.
03 / 04
Approve the reading, not just the script

Give the narrator and a fluent reviewer a short passage with a personal name, an abbreviated date, a decimal measurement and a deliberate pause. Have them listen without the script first. Can they repeat the measurement correctly? Is the name recognisable? Then compare the spoken passage with the approved text. A pleasant voice reading the wrong century is still a failed exhibit.
If a known narrator is part of the brief, agree which voice sample may be used, for which languages, for how long, and who can generate new lines. Decide what happens to old downloads if permission ends. If the team cannot staff that process, commission a recording or choose an authorised stock voice. Do not turn an archive interview into a reference sample simply because the audio file is available.
04 / 04
Move one stop through three delivery options

| Delivery | When it fits | What to prove |
|---|---|---|
| Approved audio files | The possible passages are known before opening. | The corrected track replaces the old one, including on loan devices. |
| Server-side speech | Text changes during a visit and connectivity is dependable. | A lost connection gives the visitor readable text or a stored recording, not silence. |
| Local generation | Text varies after download and the device fleet is known. | A prepared device can generate an unseen approved passage with its network disabled. |
For the third option, hand the supplier a newly reset loan phone and one used yesterday. Start the clock when a visitor taps the stop, not when a warmed-up model accepts text. Record the initial download separately from the time until intelligible speech. Repeat with a second passage and with a browser that cannot use the intended runtime. A fast repeat visit says nothing about the guest standing at the door for the first time.
For this gallery we would start with approved files and transcripts for the fixed collection, then pilot local generation only on loan devices for content that really changes. The acceptance question is simple: which passage cannot reasonably exist before the visitor arrives? If the team cannot name one, buy the recording workflow first.
Sources and editorial scope
This is Axiom Forge’s procurement and product-delivery guidance, not a guaranteed project plan. The example is illustrative. Sources checked on 15 Nov 2026.
- Zyphra: ZONOS3-Mini — Developer announcement dated 9 November: the predecessor, local runtime and execution options.
- ZONOS3-Mini technical report — Author-reported architecture and evaluation; submitted 12 November.
- Zyphra model card and files — Published weights, runtime paths, licence and language coverage.
- Webdiga: local browser and CPU checks — Independent coverage dated 12 November; not a museum deployment or a mobile listening panel.