Azure Neural TTS

Microsoft Azure

Regional Swahili and African English speech synthesis.

Pilot betaInvited workspaces · Azure South Africa North

Clear usage rates.

Characters≈ UGX 79,900per 1M characters

Introductory pilot rates · UGX

$20 per million input Unicode characters, counted once per attempt and rounded up to whole micro-USD. Budget is reserved before dispatch. Work cancelled or interrupted after dispatch may still be charged; check its receipt before retrying.

A little head start.

90 days of lower rates.

No platform subscription fee. Pay for your usage, within your approved pilot limits.

How the pilot offer works

Rates apply for 90 days from activation in an approved workspace. Rates and usage limits are confirmed before paid use.

Introductory rates are active for new requests, including existing workspaces. Historical usage keeps its original rates. Billing shows your configured rates; no automatic payments.

We’ll give at least 30 days’ notice of changes after your introductory period. Access is invite-only and subject to scheduled compute capacity. USD is the pricing base; local equivalents are approximate.

Check your workspace’s active rates →

Open in Playground View active workspace rates

Requires an invited account and a key enabled for this model.

What you can build.

Capabilities available in the Embiro pilot.

InputText

OutputAudio

Swahili voices from Kenya and Tanzania
African English voices
Regional managed inference in South Africa North
Synchronous WAV or durable queued jobs

Choose a voice.

Seven tested standard neural voices. All use the same rate; no voice cloning or custom SSML.

Zuri · sw-KE: sw-KE-ZuriNeural · female
Rafiki · sw-KE: sw-KE-RafikiNeural · male
Rehema · sw-TZ: sw-TZ-RehemaNeural · female
Daudi · sw-TZ: sw-TZ-DaudiNeural · male
Asilia · en-KE: en-KE-AsiliaNeural · female
Ezinne · en-NG: en-NG-EzinneNeural · female
Leah · en-ZA: en-ZA-LeahNeural · female

Listen to voice samples

Know the limits.

Text: Up to 1500 Unicode code points, including spaces and punctuation
Languages: en · English, sw · Swahili; voice determines language when omitted
Output: Mono PCM16 WAV at 24 kHz, up to 180 seconds; speed 1
Voices: Seven preset standard neural voices; no voice cloning or custom SSML

Full pilot service notes

Optional regional managed speech in South Africa North. Seven English/Swahili preset voices; 1500 Unicode code points; plain text only; WAV at 24 kHz, up to 180 seconds; natural speed 1. No automatic fallback or GPU warm-up. Cancelled/failed work may be charged after provider dispatch.

Requests share your workspace’s allowance, rate limit and concurrency limit. Backend capacity varies; some GPU services warm up on demand. Model access does not guarantee immediate capacity.

Make your first request.

Use a server-side workspace API key from Platform → API keys. Examples use https://api.embiro.ai/v1 and the exact pilot model ID.

API model IDazure-neural-tts

POST/v1/audio/speech

Azure TTS supports seven English/Swahili preset voices. Up to 1500 Unicode characters, WAV at 24 kHz, natural speed 1. If language is omitted it is inferred from voice. This model is optional and never automatically replaces Kokoro or Orpheus.

# Set EMBIRO_API_KEY in your shell first.
: "${EMBIRO_API_KEY:?Set your workspace API key first}"
curl --fail-with-body --silent --show-error "https://api.embiro.ai/v1/audio/speech" \
  -H "Authorization: Bearer $EMBIRO_API_KEY" \
  -H "Content-Type: application/json" \
  --output speech.wav \
  --data-binary @- <<'JSON'
{
  "model": "azure-neural-tts",
  "input": "Habari yako rafiki. Karibu Embiro Labs.",
  "voice": "sw-KE-ZuriNeural",
  "language": "sw",
  "response_format": "wav"
}
JSON

What comes back

A WAV response body. The example saves speech.wav.

Cold starts, errors and safe retries

Use the Queued job example when a request can wait for compute. Save its ID and poll that job.
For 401/403, check the key and its model permissions. For 400, check the request and format limits.
Respect Retry-After on 429/503. A timeout after dispatch can be ambiguous; check request logs before submitting again.
These examples do not automatically retry paid work. A job retry must reuse the saved idempotency key and identical payload.

Open API reference

Under the hood.

Publisher specifications. These do not override the Embiro service limits above.

Published context: Not specified
Parameters: Not disclosed
Model size: Managed

Managed standard neural speech in South Africa North. Seven preset voices; 1500 Unicode characters, WAV at 24 kHz and natural speed 1. Provider model size and version are not disclosed. No voice cloning or customer SSML.

Read the publisher’s model card

No verified benchmark scores are listed in this catalogue for this model.