Thanks — this is a great question and a very common challenge when targeting emerging markets. You mentioned OpenAI’s realtime voices sound too American; below I’ll acknowledge that and give practical, actionable options you can test with LiveKit to get more authentic local accents in English and French. Acknowledgement • You want reproductions of Indian, Kenyan, Nigerian, Moroccan, and Senegalese accents (English and French) and find default US-style voices unsuitable. Top approaches (what you can do right away) • Try TTS providers that specialize in multilingual and expressive voices (better regional prosody than US-centric voices). • Use descriptive prompting and voice-personality controls where available to steer accent, plus SSML/phonetic spelling to correct pronunciations that sound “off.” • If no native regional voice exists, emulate accents via prosody + phonetic tweaks or by cloning recordings of target speakers where permitted. Recommended providers and why they fit • Rime (recommended for Indian English) ◦ Strong multilingual models and good handling of Indian English pronunciation and code-switching. ◦ Use Rime voices via LiveKit Inference for lower-friction integration. • Cartesia ◦ Fast, high-quality TTS with strong prosody control. Good for tuning pitch/rate to match regional intonation. ◦ Use SSML tags like <prosody> or <phoneme> to nudge pronunciation toward local variants. • Hume (best for expressive/emulation) ◦ Lets you describe the speaker (e.g., “a calm West African male speaker with a soft French accent”) and tune pace/energy, which works well when a specific labeled regional voice is not available. ◦ Useful for Moroccan French or West African French/English where a dedicated voice may be missing. • ElevenLabs (custom cloning path) ◦ Default voices skew US/UK, but if you can create custom voice clones from local speakers you can achieve highly authentic accents. Note this requires voice samples and ElevenLabs account-level tooling. Practical techniques to improve perceived accent quality 1. SSML + phonetic spelling ◦ Use

<prosody>

to slow or lower pitch and

<phoneme>

/ phonetic spellings to force local pronunciations for tricky words or names. 1. Descriptive prompting ◦ For Hume-style models, include short natural-language descriptions: e.g., “Speak like a Kenyan news presenter: measured pace, neutral West African intonation.” 1. Voice cloning (when legal/ethical) ◦ If you have consented recordings of local speakers, clone a voice (ElevenLabs or similar) to get authentic accent and timbre. 1. Pipeline tuning ◦ Combine a locally robust STT (if you need recognition) and then feed cleaned transcripts to the TTS layer tuned for accent—this helps maintain natural prosody and reduces mismatches. 1. A/B test multiple providers ◦ Run short script comparisons across Rime, Cartesia, Hume, and any custom clones to pick the best fit per region/language. Region-specific recommendations • India (English): Rime first, Cartesia second. Rime handles code-switching and Indian phonetics best. • Kenya / Nigeria (English): Hume or Cartesia with descriptive prompts and SSML tweaks; consider custom cloning if you need a very local sound. • Morocco / Senegal (French): Hume with descriptive voice prompts or Cartesia with SSML. If you need a Maghrebi French or West African French timbre, emulate via prosody + phonetic adjustments. Next step (so I can give targeted help) • Which specific region and language pair do you want to prioritize first (e.g., Nigerian English, Moroccan French)? • Do you already have any short sample lines or a target speaker recording we can use to test emulation? If yes, tell me one short phrase to use for comparisons. • Are you using LiveKit Inference only, or are you able to use LiveKit plugins (Hume plugin, Rime via Inference, etc.)? If you tell me the first region/language and whether you can run plugins or only inference, I’ll provide a short test plan you can run and example SSML/prompt text to try with two providers so you can compare quickly. Sources: Build a multilingual voice agent that automatically switches languages | Li... | Rime TTS | LiveKit Documentation | Cartesia TTS | LiveKit Documentation | Hume TTS plugin guide | LiveKit Documentation | Text-to-speech (TTS) models overview | LiveKit Documentation