Speech-to-text

POST /v1/stt transcribes recorded audio with Waridi Sikia (sikia-1). It is built for the way people in East Africa actually talk: regional accents, Swahili–English code-switching and everyday Sheng.

For live audio, use the Streaming API instead.

Request

Upload a file as multipart/form-data, or send JSON with an audio_url we can fetch.

from waridi import Waridi

client = Waridi()

with open("simu.wav", "rb") as f:
    result = client.stt.transcribe(
        model="sikia-1",
        file=f,
        language="sw-TZ",
        code_switching=True,
        timestamps="word",
    )

for segment in result.segments:
    print(segment.start_ms, segment.language, segment.text)

import { Waridi } from "@waridi/sdk";
import { createReadStream } from "node:fs";

const client = new Waridi();

const result = await client.stt.transcribe({
  model: "sikia-1",
  file: createReadStream("simu.wav"),
  language: "sw-TZ",
  codeSwitching: true,
  timestamps: "word",
});

for (const segment of result.segments) {
  console.log(segment.startMs, segment.language, segment.text);
}

curl https://api.waridi.io/v1/stt \
  -H "Authorization: Bearer $WARIDI_API_KEY" \
  -F model=sikia-1 \
  -F language=sw-TZ \
  -F code_switching=true \
  -F timestamps=word \
  -F file=@simu.wav

Parameters

Field	Type	Required	Description
model	string	yes	sikia-1
file	file	one of	Audio file up to 200 MB and 2 hours
audio_url	string	one of	HTTPS URL to an audio file. Use instead of file
language	string	no	sw-TZ (default), sw-KE or auto
code_switching	boolean	no	Default true. Tags each segment with the language spoken
timestamps	string	no	none, segment (default) or word
diarization	boolean	no	Beta. Labels speakers as spk_0, spk_1 and so on
vocabulary	string[]	no	Up to 100 words or phrases to bias toward, such as product names

Accepted input formats: WAV, MP3, M4A, Ogg/Opus, FLAC and WebM. Telephony audio at 8 kHz μ-law is accepted as WAV or raw with encoding=mulaw.

Response

{
  "id": "stt_01J9Z6R2M1",
  "model": "sikia-1",
  "language": "sw-TZ",
  "duration_ms": 6420,
  "text": "Nataka kuangalia salio langu, na pia ku-check kama malipo yamefika.",
  "segments": [
    {
      "start_ms": 0,
      "end_ms": 2310,
      "language": "sw",
      "text": "Nataka kuangalia salio langu,",
      "speaker": "spk_0"
    },
    {
      "start_ms": 2310,
      "end_ms": 6420,
      "language": "sw-en",
      "text": "na pia ku-check kama malipo yamefika.",
      "speaker": "spk_0"
    }
  ]
}

Field	Description
text	Full transcript as one string
segments[].language	sw, en, or sw-en when a segment mixes both
segments[].words	Present when timestamps is word. Each word has start_ms, end_ms and confidence
duration_ms	Length of the input audio, used for billing

Code-switching

We do not translate English words into Swahili or drop them. If a caller says "nime-transfer pesa jana", you get exactly that. Sikia keeps Swahili verb prefixes attached to English stems, which is how people write it too.

Sheng support is experimental. Expect lower accuracy on fast Nairobi slang than on standard Swahili, and tell us where it falls short at hello@waridi.io.

Accuracy

sikia-1 is a preview model. We are building a public Swahili ASR benchmark and will publish results with methodology when it is ready, not before.