Bridge by Wekoya / Docs / Media

Voice

Updated 8 November 2026

Turn calls, voice notes and lessons into text, and text into natural speech, on the same key and balance. The routes are OpenAI's, so the OpenAI SDK's audio.transcriptions and audio.speech work as they are: point the client at Bridge and use nova-3 to listen and aura-2 to speak, both by Deepgram, or Google's Gemini models below.

import os
from openai import OpenAI

bridge = OpenAI(api_key=os.environ["WEKOYA_API_KEY"], base_url="https://wekoya-api.wekoya.tech/v1")

# Speech to text: a recording up to 25 MB and 60 minutes.
with open("call.mp3", "rb") as audio:
    transcript = bridge.audio.transcriptions.create(model="nova-3", file=audio)
print(transcript.text)

# Text to speech: up to 2,000 characters a request.
speech = bridge.audio.speech.create(
    model="aura-2",
    voice="thalia",
    input="Your transfer of fifty thousand naira was received.",
)
speech.write_to_file("reply.mp3")

import fs from "node:fs"
import OpenAI from "openai"

const bridge = new OpenAI({ apiKey: process.env.WEKOYA_API_KEY, baseURL: "https://wekoya-api.wekoya.tech/v1" })

const transcript = await bridge.audio.transcriptions.create({
  model: "nova-3",
  file: fs.createReadStream("call.mp3"),
  response_format: "verbose_json",          // segments with start and end times
  timestamp_granularities: ["word", "segment"],
})

const speech = await bridge.audio.speech.create({ model: "aura-2", voice: "thalia", input: "Welcome back.", response_format: "wav" })
fs.writeFileSync("welcome.wav", Buffer.from(await speech.arrayBuffer()))

A transcription answers with its cost, like a chat reply. response_format can be json, text, verbose_json (segments, and words with timestamp_granularities), srt or vtt; for text formats and for speech, the cost is in the X-Bridge-Cost header.

{
  "text": "Your transfer of 50,000 naira was received on October 9.",
  "usage": { "type": "duration", "seconds": 7 },
  "wekoya": { "cost": 0.89, "currency": "NGN", "seconds": 7 }
}

# Who said what: diarize labels each segment and word with a speaker.
curl https://wekoya-api.wekoya.tech/v1/audio/transcriptions \
  -H "Authorization: Bearer $WEKOYA_API_KEY" \
  -F file=@meeting.m4a -F model=nova-3 -F diarize=true -F response_format=verbose_json

# Subtitles, ready for a video player:
curl https://wekoya-api.wekoya.tech/v1/audio/transcriptions \
  -H "Authorization: Bearer $WEKOYA_API_KEY" \
  -F file=@lesson.mp3 -F model=nova-3 -F response_format=srt

# The voices, filtered by language:
curl "https://wekoya-api.wekoya.tech/v1/audio/voices?language=en" -H "Authorization: Bearer $WEKOYA_API_KEY"

Languages and voices

language is English (en) unless you set it. Use English for English spoken with any accent, an ISO code such as fr, es, pt or de for another language, or multi when a recording moves between languages. For Hausa, use gemini-3.5-transcribe. Yoruba, Igbo and Nigerian Pidgin are not supported yet, and a request for them is refused with nothing charged. Aura voices speak English, Spanish, French, German, Italian, Dutch and Japanese: name one by its short name (thalia) or full id (aura-2-thalia-en), and list them all with GET /v1/audio/voices (add ?model=gemini-3.8-flash-tts for Gemini's, each with its style). For a dialogue, pass two speakers ([{"name": "Ada", "voice": "Kore"}, {"name": "Tunde", "voice": "Puck"}]) and write the input as Ada: … lines. OpenAI's own voice names such as alloy and its instructions are not available: these voices would read instructions aloud, so choose a voice whose style fits. Speed control and audio translation are not offered. For audio as it is spoken, see Live transcription.

Everything that applies to chat applies here: key scopes and model allow-lists, end users and their budgets, and the same errors. A request that fails costs nothing. Voice runs on your Wekoya balance only. Recordings are not used to train any model. Voice models are left out of GET /v1/models so chat pickers never offer them; list them with GET /v1/models?type=audio.