Voice cloning

Voice cloning lets an event’s interpretation use a voice created from a speaker’s own recording. Events use Fidelity with Gemini speech synthesis; the same saved voice can be selected for several interpreted languages.

Create a voice#

1. Sign in and select the organization that owns your event.
2. Open Voices. You need to be the organization’s owner or an admin to create or delete voices.
3. Give the voice a name you will recognize in the language settings.
4. Record or upload 10–30 seconds of the speaker’s natural speech. Use a clear recording of one speaker, without music or other people talking over them.
5. Record the same speaker reading the consent statement shown on the page, word for word. Confirm that the speaker agrees to their voice being replicated.
6. Create the voice. Google checks the recordings and consent before returning the saved voice.

The browser converts the clips to mono, 24 kHz, 16-bit WAV. Each organization can keep up to 20 cloned voices.

Test a voice#

Select Test next to a saved voice in Voices to hear it before using it at an event. Every member of the organization can test its voices.

1. Choose the language to read in. The voice reads the text the way an event’s interpretation into that language would.
2. Select an example — a welcome, an announcement or a short talk — or write your own text of up to 500 characters. Examples are available in common languages; for other languages, write your own text in that language.
3. Select Play. The reading plays below the text and can be replayed.

Each test reading is generated by Google. An organization can make 60 test readings an hour.

Use it at an event#

Open a language’s Settings → AI voice, select your saved voice and save the settings. The voice must belong to the event’s organization. Select it separately for each language that should use it; changes take effect when that language channel next starts.

Cloned voices work with Fidelity while Gemini is the speech provider. They do not change phone-call voices: phone calls continue to use gpt-realtime-translate. The public demo’s voice picker offers stock voices; create an event to try your own saved voice.

Recordings and deletion#

The sample and consent recordings are sent to Google to create the voice. Spiiksi keeps the voice’s name, provider ID and consent date; it does not retain the uploaded clips. Google manages the stored voice profile. See Google’s voice-replication documentation for provider requirements and retention.

Delete a voice from Voices when you no longer need it. Successful deletion removes it from Google and Spiiksi and clears its selection from the organization’s language channels, which return to their stock voice.

If creation fails#

Read the explanation shown beneath the form. Check that both recordings are from the same speaker and that the consent statement matches the displayed text. If needed, record both clips again in a natural voice.

Google sometimes returns the generic error “Error translating server response to JSON.” Spiiksi shows a more specific provider explanation when one is available. A generic error alone does not establish whether the recording, consent check or provider service caused the failure. If it persists, contact support with the error text and time of the attempt.