Change the speaker voice
The speaker parameter selects the voice for the generated speech. Speaker names are lowercase and case-sensitive (shubh, not Shubh).
Hear the voices before you choose: the Voices page has audio samples for every speaker, and you can try your own text in the playground.
Choose a speaker
Each model has its own set of speaker IDs. A speaker from one model is rejected on the other.
The language tag in a v4 persona ID is not a restriction. Any supported language_code works with any persona.
Bulbul v3 speakers
Example
This example uses a bulbul:v4-flash persona. For bulbul:v3, set model to "bulbul:v3" and speaker to a v3 name such as "priya".
On WebSocket, set speaker in the configure message. See the WebSocket API.
Common errors
The message lists every valid speaker for the model in the request. On WebSocket, a bad speaker returns a JSON error message with code 422.
Recommended speakers (Bulbul v3)
Speakers are rated by Critical Error Rate (CER), the percentage of words with critical pronunciation errors. Lower is more reliable. These ratings apply to bulbul:v3 speakers.
varun has a very low CER (0.06%) but a deep, dramatic villain or suspense character. Do not use it as a neutral default. Reserve it for thriller, drama or suspense content.
Recommended speakers by language
Speakers do not perform equally in every language. Prefer these picks for each language:
Top picks: priya and ishita (best female, across Hindi, Telugu, Kannada, Tamil, Marathi, Gujarati and English) · mani (best male overall, Punjabi) · shubh (best male for hi, te, kn, od, ml) · ratan (best male for en, te, kn, ta, mr, gu).