Speech to text

Tapra-Voice

Batch Tunisian speech to text, with the code-switching intact.

Tapra-Voice is the offline half of the speech family. Upload a call recording, a voice note, or a video soundtrack and get one transcript: Derja in Arabic script, with every French and English word kept in Latin script, exactly where the speaker said it. It is the model behind Tapra Subtitles, Video Translation, and Audio Search.

See it run API reference

0.150 TND / min
word + segment times
SRT / VTT ready

POST /v1/audio/transcriptions

tapra-voice-v1

Audio

talk_show_clip.mp4 · 58 s

Segment 16.48 – 18.24 s

اما she is frustrated على خاطر he is shorter than her

Live output

Real output

Recorded from the live model.

Every output on this page came from tapra-voice-v1 on 8 October 2026 and is shown unedited, mistakes included. Each live response also carries its exact charge in x-tapra-cost-tnd.

A 26-second voice note that moves between Derja, French, and English. Press play: the segment being spoken lights up.

0:00 / 0:00

Audio · 26 s

Transcript · timestamps=segment

1.44نخمم نعمل une combinaison ما بين les deux
3.68modèles tomacans اللي هو capable يجاوب على
7.44les clients كيما مثلا في
9.92confirmation متاع
11.20les ordres والا تعمل
13.44marketing
14.56لهنا le modèle لازمو يكون قادر يعمل switch
19.04between languages ويجاوب زادا بنفس الفاصن

Unedited model output. It mishears one product name ("tomacans") and writes façon as "الفاصن". Every French and English word otherwise stays in Latin script.

What it does

Capabilities you can build on today.

One request, three languages

The decoder runs Darija, French, and English together, so "نكلمو ال client بعد ال réunion" stays a single sentence instead of three broken ones.

Word-level timestamps

timestamps=word returns t0_ms/t1_ms per token for karaoke captions; timestamps=segment returns cue-ready blocks for SRT and VTT exports.

Vocabulary steering

Pass up to 500 characters of expected words in prompt and tune context_boost (0–2) to hold brand names, street names, and product terms steady.

Metered by audio, not by wall clock

You are billed on the audio milliseconds you send at 0.150 TND / min. No idle GPU, no minimum session, no cold-start penalty.

Integrate

The contract, in one place.

Model id: tapra-voice-v1
Endpoint: POST /v1/audio/transcriptions
Input formats: WAV, MP3, M4A, OGG, WebM
Language hint: language, default ar-AR (code-switching stays on)
Timestamps: segment | word
Speakers: diarize=true adds speaker labels at no extra charge
Punctuation: punctuation=true restores it; off by default
Prompt: Optional, max 500 characters
Context boost: 0 – 2, default 1
Upload limit: 100 MB per file
Response: text, segments, words, duration, usage, x-tapra-cost-tnd

Transcribe a filebash

curl -X POST "$TAPRA_API_URL/v1/audio/transcriptions" \
  -H "Authorization: Bearer $TAPRA_API_KEY" \
  -F "file=@customer_call.wav" \
  -F "model=tapra-voice-v1" \
  -F "language=ar-AR" \
  -F "timestamps=word"

Comparison

How Tapra-Voice differs.

Compared with a generic multi-language cloud ASR, which is usually a set of monolingual decoders behind one endpoint.

Tapra-Voice compared with a generic alternative
Capability	Tapra-Voice	Generic alternative
Darija + French + English in one pass	Single decoder, switches mid-sentence	Language listed per request; the switch is flattened or dropped
French and English words	Written in Latin script, in place, inside the Arabic line	Transliterated into Arabic letters or dropped
Timestamps	Segment and word level on every plan	Usually segment-only, word timing behind a premium tier
Compute	Commodity CPU, no GPU cold start	Shared GPU pools billed while idle
Billing	0.150 TND / min (≈ $0.050), metered exactly	Per-minute rounding with regional markups
Currency	Charged in TND, top-ups by local transfer	International card required

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

0.150 TND / min

Files in, transcript out. Darija, French, and English are decoded in the same pass, with timestamps ready for subtitles.

Start with 2 TND free See plans

Rate: $0.050 per audio minute
In dinars: 0.150 TND per audio minute
Hourly equivalent: $3.00 per hour
Metered unit: audio_ms (exact audio duration)
Minimum charge: None — no idle session fee

Streaming transcription is a separate model, billed at 0.300 TND / min: Tapra-Voice-Realtime.

Questions

What builders ask first.

How long can a single file be?

Long recordings are expected: the endpoint streams the upload to the model host and returns one segment list. Very long media is better split by the Subtitles service, which handles chunking and cue merging for you.

Does it translate while transcribing?

No. Tapra-Voice only transcribes. Pair it with Tapra-Translate to get subtitles in another language, or use the Video Translation service which chains both.