SAUTI — Swahili Voice AI

Speak Swahili to an AI assistant

Record a message in Swahili (or type below). The AI understands and responds in Swahili with voice.

Pipeline: Your voice → ASR (Whisper) → LLM (Llama 3.1) → TTS (MMS) → AI voice

Record your message (Swahili)

Or type in Swahili:

Type Swahili

Conversation

AI Response

Latency

Swahili Speech Recognition

Upload or record Swahili audio. Our fine-tuned Whisper model transcribes it with 13.5% WER — trained on 89 hours of Swahili speech.

Model: Finiflowlabs/sauti-asr-v1

Upload or record Swahili audio

Clone Any Voice

Upload 6-30 seconds of someone's voice, then type text to hear the AI speak in that voice. Uses XTTS v2 zero-shot cloning.

Note: Voice cloning requires GPU hardware for acceptable speed. On CPU, synthesis takes ~30-60s per utterance. XTTS v2 works best with English text; Swahili uses cross-lingual transfer.

Reference audio (6-30 seconds)

Text to synthesize

Cloned voice output

Swahili Text-to-Speech

Type Swahili text and hear it spoken using Meta's MMS-TTS model.

Model: facebook/mms-tts-swh

Enter Swahili text

Try these examples

Synthesized audio

English ↔ Kiswahili Translation

Speak or type in one language, hear and read the translation in the other.

Pipeline: Your voice → ASR (Whisper) → MT (NLLB-200) → TTS (MMS) → Translated voice

Note: First use loads the translation model (~60s on CPU, ~10s on GPU).

Source language

Target language

Record or upload audio

Or type text to translate:

Enter text

Source transcript

Translation