Text to speech

Tapra-Speak

A Tunisian voice that sounds like it grew up here.

Tapra-Speak turns Arabic script or Arabizi into speech with Tunisian prosody instead of MSA recitation. It is the model behind Tapra Voice Studio: type a script, pick a voice, download 24 kHz WAV. On a call flow it answers customers in the accent they called in with.

See it run API reference

0.200 TND / 1k chars
24 kHz WAV
clone 1.25×

POST /v1/audio/speech

tapra-speak-v2.1

input · instruct

أهلا خويا، نتلاقاو غدوة après le travail؟

Audio · 24 kHz mono WAV · 3.25 s

instruct: "Male, Young Adult, Moderate Pitch"

Live output · num_step 16 · speed 1.0

Real output

Recorded from the live model.

Every output on this page came from tapra-speak-v2.1 on 8 October 2026 and is shown unedited, mistakes included. Each live response also carries its exact charge in x-tapra-cost-tnd.

Five voices, five everyday scripts. Each voice is designed from a short instruct string, with no reference audio. Press play.

input · instruct

أهلا خويا، نتلاقاو غدوة après le travail؟

Audio · 24 kHz mono WAV · 3.25 s

instruct: "Male, Young Adult, Moderate Pitch"

0:00 / 0:00

num_step 16 · speed 1.0

What it does

Capabilities you can build on today.

Tunisian prosody, not MSA recitation

Numbers, greetings, and filler words land the way a Tunisian speaker says them, which is what makes a voicebot sound local rather than dubbed.

Voice design and pacing

instruct designs a voice from a short description such as "Female, Middle-aged, Low Pitch". speed (0.7 – 1.3) sets the pace, and duration can pin the clip to a target length.

Instant voice clone

Send a short reference clip and speak in that voice. Clone requests are billed at 1.25× the character rate because the speaker embedding is the premium product.

WAV or JSON, your pick

response_format: "wav" streams 24 kHz audio bytes; "json" returns the same audio base64-encoded with duration and usage attached.

Integrate

The contract, in one place.

Model id: tapra-speak-v2.1
Endpoint: POST /v1/audio/speech
Output: 24 kHz WAV (base64 in JSON, or raw audio/wav)
Character limit: 2,000 characters per request
Voice design: instruct, e.g. "Male, Young Adult, Low Pitch"; ignored with ref_audio
Pacing: speed 0.7 – 1.3, optional duration 0.2 – 120 s
Arabizi: convert_arabizi=true (default) converts Latin-typed Derja first
Voice cloning: Instant clone from a reference clip, billed at 1.25×
Billing unit: Input characters, 0.00295 TND minimum per request
Response extras: duration, processing_ms, usage, x-tapra-cost-tnd

Generate a voiceoverbash

curl -X POST "$TAPRA_API_URL/v1/audio/speech" \
  -H "Authorization: Bearer $TAPRA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tapra-speak-v2.1",
    "input": "أهلا بيك في Tapra، نجم نعاونك في شنوّة؟",
    "response_format": "wav"
  }' --output greeting.wav

Comparison

How Tapra-Speak differs.

Compared with MSA-first TTS APIs that read Arabic as fusḥa.

Tapra-Speak compared with a generic alternative
Capability	Tapra-Speak	Generic alternative
Dialect	Tunisian Derja prosody and lexicon	MSA recitation; dialect voices are licensed add-ons
Arabizi input	Converted by Tapra-Script inside the same request	Requires a separate converter call
French and English words	Spoken in place inside the Arabic sentence	Often read letter by letter or with a heavy accent
Cloning	Instant clone at 1.25× the character rate	Enterprise contracts, per-seat or per-clone fees
Output ownership	WAV bytes returned to you, downloadable as WAV or MP3	Often a hosted URL with retention limits
Cost	0.200 TND / 1k chars (≈ $0.067), $0.001 minimum	$0.015+ per 1,000 characters with GPU cold starts

Pricing

Metered, never subscribed.

You pay for exactly what you send. The same prepaid balance covers all five Tapra models, billed in TND.

Headline rate

0.200 TND / 1k chars

Conversational Derja synthesis for voicebots, IVR, and accessible media — with voice design, pacing control, and an instant voice clone.

Start with 2 TND free See plans

Rate: $0.067 per 1,000 characters
In bulk: $66.72 per 1M characters
In dinars: 200.00 TND per 1M characters
Metered unit: characters (input length)
Minimum charge: $0.001 per request

Instant voice clone requests are billed at 1.25× the Speak character rate.

Questions

What builders ask first.

Can I hear a voice before cloning one?

Yes. Voice Studio lists the built-in Tunisian voices and plays a sample. Cloning is a separate request that carries the reference clip.

Why is cloning billed differently?

Clone requests are 1.25× the Speak character rate. The reference clip already consumed storage and the speaker embedding is the part of the product with real marginal cost.