Text-to-Speech Integrations
Text-to-speech turns your agent responses into spoken audio. Voice quality shapes how callers perceive your brand. Flat, robotic speech kills trust while natural, expressive voices build it. Bolna integrates with the fastest TTS providers so responses sound human and arrive without awkward pauses.
Amazon Polly
Amazon Polly gives your Bolna voice agent low-cost text-to-speech with generative, neural and standard engines across a broad range of locales.
Azure Text-to-Speech
Azure Text-to-Speech gives your Bolna voice agent Microsoft's neural voices in many languages, covered by Bolna's flat per-minute rate on the neural model.
Cartesia
Cartesia gives your Bolna voice agent fast Sonic text-to-speech in 42 languages, including Hindi, with voice cloning and your own API key.
Custom TTS
Custom TTS lets your Bolna voice agent speak through your own OpenAI-compatible text-to-speech server, such as vLLM-Omni or an open-weight model, with your key.
Deepgram Aura
Deepgram Aura gives your Bolna voice agent fast, accurate English text-to-speech with the aura-2 model, and pairs with Deepgram transcription on one account.
ElevenLabs
ElevenLabs is Bolna's default voice. It gives your agent natural speech, Indian and American voices, voice cloning and models covering up to 74 languages.
Kalpa Labs
Kalpa Labs gives your Bolna voice agent text-to-speech that speaks English, Hindi and code-switched Hinglish from one model, with no language setting needed.
Maya
Maya gives your Bolna voice agent native-sounding text-to-speech in 10 Indian languages plus Indian English, with two voices and mid-call language switching.
Rime
Rime gives your Bolna voice agent natural, conversational English speech through its Arcana and Mist v2 models, with your own Rime account.
Sarvam AI (TTS)
Sarvam Bulbul v3 gives your Bolna voice agent native-sounding speech in Hindi, Tamil, Telugu and 8 more languages, with 16 voices to choose from.
Smallest AI
Smallest AI gives your Bolna voice agent low-latency Lightning text-to-speech, with 149 Pro voices in Indian, British and American accents.
Soniox (TTS)
Soniox tts-rt-v2 gives your Bolna voice agent real-time streaming speech in 63 languages from one model, with 200 named voices and voice cloning.
Text-to-speech providers on Bolna compared
| Provider | Best for | Languages | Notes |
|---|---|---|---|
| Sarvam (bulbul:v3) | Indian languages | 10 Indian languages plus English (India) | 16 native voices; language must be set in provider_config; included in the flat rate |
| ElevenLabs | Natural English voices | eleven_v3_conversational and eleven_v4_turbo cover 74 languages | Bolna default; eleven_turbo_v2_5 recommended, eleven_flash_v2_5 fastest; supports voice cloning |
| Cartesia | Lowest latency English | Sonic models cover 42 languages, including Hindi | sonic-3.5 recommended for production; supports voice cloning |
| Rime | Natural conversational English | English and Spanish | Arcana and Mist v2 models; good prosody for dialogue |
| Smallest AI | Ultra-low latency | Voices with Indian, British and American accents | lightning_v3.1_pro model with 149 Pro voices |
| Maya (Maya 2 Native) | Indian languages | 10 Indian languages plus Indian English, or auto-detect | Two voices cover every language; switches language mid-call without reconnecting; included in the flat rate |
| Kalpa Labs (kalpa-tts-1) | English and Hinglish | English and Hindi, including code-switched Hinglish | One model with no language setting |
| Soniox (tts-rt-v2) | One provider for STT and TTS | 63 languages, including 10 Indian languages | 200 named voices; voice cloning; same Soniox key as the transcriber |
| Azure Text-to-Speech | Multilingual and enterprise deployments | Over 140 languages | neural model is included in the flat rate |
| Amazon Polly | Cost-sensitive workloads | Broad range of locales | generative, neural and standard engines; set the language in provider_config |
| Deepgram Aura | English on one Deepgram account | English | aura-2 model; pairs with Deepgram transcription |
| Custom TTS | Your own or open-weight voice model | Depends on your model | Any OpenAI-compatible speech endpoint, such as vLLM-Omni |
For the full guide, see Choosing a text-to-speech provider in the Bolna docs.
Frequently asked questions
Cartesia has the lowest latency for English, with very fast time to first audio. Within ElevenLabs, eleven_flash_v2_5 is the fastest model, and Smallest AI is built for real-time output. A buffer_size of 100 to 150 characters also gets the first word out sooner.
No. Sarvam Bulbul text-to-speech covers 11 languages, fewer than the Saaras transcriber. Assamese, for example, can be transcribed but Bulbul has no Assamese voice, so pair it with another synthesizer.
Yes. Upload a 1 to 2 minute sample of a voice you have the rights to use, and host the clone on ElevenLabs or Cartesia.
ElevenLabs Flash, Turbo, Multilingual v2 and v3 conversational models, Sarvam bulbul:v3, Cartesia sonic-3, sonic-3.5 and sonic-3.6, Azure neural voices, and Maya. Other voice models are billed at usage-based rates.
Always set stream to true. Keep buffer_size between 100 and 250 characters, and match audio_format to your telephony provider.
Last updated