Arabic TTS pricing in 2026 looks deceptively simple on a vendor's pricing page (some number of USD per million characters), and turns into a thicket the moment you scale past a proof-of-concept. This post is a tour through the actual axes that matter, the line items that hide most of the cost, and how to compare vendors when their pricing pages don't expose the same primitives.
We'll quote rough market-rate ranges where useful, but we won't quote specific vendor prices; the rates change quarterly. The framework below is designed to outlast the specific numbers.
The five axes of TTS pricing
1. Per-character vs. per-second
The two dominant pricing units are per-character (counted on the input text) and per-second (counted on the output audio duration). The ratio between them depends on the language: Arabic averages roughly 14 characters per second of speech, English averages 12-14, French 13-16. So pricing identical workloads under the two units produces similar dollar figures, but the predictability is different.
Per-character pricing is predictable from the script. You know the input length before rendering, so you can budget exactly. Per-second pricing requires either rendering a sample first or estimating based on language-specific speech rates, which gets messy when scripts have variable speaking-rate intent (audiobooks slow, podcasts fast).
For a buyer building budget projections, per-character is easier to reason about. The major Arabic TTS vendors all price per-character. The major English TTS vendors are split between per-character and per-second.
2. Streaming vs. batch tier
Most vendors price streaming TTS higher than batch TTS, even when the underlying model is the same. The rationale is that streaming carries operational overhead (WebSocket fleet, low-latency inference scheduling, mid-stream cancel handling) that batch doesn't. Premium ranges 1.5-3x the batch rate at the volumes we see.
The rule of thumb: if your use case is real-time (voice agents, interactive demos, live narration), pay the streaming premium. If your use case is offline (audiobooks, video, e-learning), pay the batch rate even if the streaming endpoint would technically work.
3. Voice catalogue vs. voice cloning
Catalogue voices are billed per-character on output. Voice cloning has two sub-axes:
- Cloning fee: a flat rate (or sometimes a tiered rate) per cloned voice. Charged once per clone, regardless of how much output the cloned voice generates. Market rates in 2026: USD 50-500 per clone depending on the vendor and the cloning quality tier.
- Per-character on cloned-voice output: most vendors bill cloned-voice generations at a premium over catalogue voices, typically 1.2-2x. Some vendors don't charge a premium and recover the cost through the cloning fee; others charge both.
For a customer with one cloned voice and high output volume (a podcast running for years on a cloned host voice), the cloning fee is amortised quickly and the per-character premium dominates. For a customer with many cloned voices and low output per voice (a dubbing studio cloning each character in a series), the cloning fee dominates.
4. Free tier vs. paid tier voice quality
Some vendors ship the same model across the free and paid tiers; the free tier is a usage limit on the same product. Others ship a degraded model on the free tier (lower sample rate, fewer voices, no streaming) so the paid tier feels like a step change.
This matters for evaluation. If you benchmark a vendor on the free tier and the paid tier ships a different model, your benchmark is invalid. Ask explicitly: "is the free-tier model the same model as the paid-tier model?". A vendor that gives a clear yes is a vendor whose free tier is a useful evaluation surface.
Hakim's free tier ships the same model as the paid tier, with a 10,000 characters/month usage limit. The free tier covers a chapter of evaluation; the paid tier kicks in for production volumes.
5. Dialect surcharges (and other per-feature charges)
Some Arabic TTS vendors price MSA at the base rate and charge a surcharge for regional dialects. The reasoning given is that dialects required additional training data acquisition; the actual reasoning is usually that the dialect tier is a value-tier upsell. Surcharge ranges from 10-50% over the MSA base rate.
Other per-feature charges that show up in contract pricing (rare on public pages):
- SSML usage premium for technical content with extensive prosody control.
- Higher sample-rate output (48 kHz studio masters vs. 22 kHz default).
- Long-context generation (audiobooks longer than 100,000 characters per batch) priced as a separate SKU.
- Custom voice tuning (fine-tuning a catalogue voice on your brand's prosody profile) as a one-time engagement fee.
A buyer should ask the vendor for the full list of per-feature charges, not just the headline per-character price.
What it costs to ship common Arabic TTS workloads
Rough market-rate figures for 2026, useful for budget projections:
- A 1-hour Arabic podcast episode: ~50,000 characters of script, ~USD 5-15 in batch TTS. Compare with USD 200-500 for a voice actor.
- An 8-hour Arabic audiobook: ~400,000 characters of script, ~USD 40-100 in batch TTS. Compare with USD 2,000-5,000 for a studio booking with a voice actor.
- A 12-module e-learning course (6 hours of audio): ~300,000 characters of script, ~USD 30-75 in batch TTS, with a multi-dialect render adding the cost of additional dialect renders (each render runs the full script again). Compare with USD 1,500-3,000 for an instructor recording.
- A voice agent serving 10,000 calls/month at average 2 minutes per call: ~30 million characters/month of streaming TTS output, ~USD 600-1,500 in streaming TTS at typical volume-discounted rates. Compare with USD 30,000-60,000 for human agents handling the same volume (loaded labour cost).
- A real-time dubbing pipeline rendering 4 hours of dub per day: ~720,000 characters/day batch TTS, ~USD 75-180/day, ~USD 2,200-5,500/month.
These are TTS-only figures; for full workflows (voice cloning, dubbing pipelines with translation overhead, voice agents with STT and LLM costs) the TTS line is typically 20-40% of the total spend.
The free-tier traps
Three patterns to watch for on free tiers:
- Free-tier voices that don't ship to production: some vendors gate the production-quality voices behind the paid tier. The free tier ships an "evaluation voice" that's useful for trying the API but not for producing actual content. Test the same voice on free and paid before benchmarking.
- Free-tier rate limits that don't reflect production: a vendor whose free tier rate-limits at 1 request/second can't be benchmarked for streaming voice agent latency, because streaming requires sustained throughput. Use the paid trial (most vendors offer one) for latency benchmarking.
- Free-tier output watermarks: a vendor whose free-tier output carries an audible watermark or a "produced with X" announcement isn't shipping the same product as the paid tier. Production deployment with the free tier is operationally unworkable.
Volume discounting
At scale, every TTS vendor has a volume-discount structure that's not on the public pricing page. Typical patterns:
- Tiered per-character pricing: USD X per million for the first N million, USD Y for the next, etc. The break points are typically at 1M, 10M, 100M, 1B characters/month.
- Annual commit pricing: a fixed annual fee for a guaranteed character allowance, with overage at a higher rate. Useful for customers with predictable monthly volumes.
- Reserved capacity: dedicated inference capacity that's always available, charged as a flat monthly fee regardless of usage. Useful for voice agents where unmet TTFB targets are worse than over-provisioning.
- Multi-year discounts: 2-3 year commitments for an additional 10-25% discount over the annual rate.
A buyer at production volume (10M+ characters/month) should always negotiate; the public price is rarely what production customers actually pay.
Hidden costs that show up at scale
Five line items that don't appear on the per-character price but matter at production volume:
- Egress bandwidth: streaming TTS produces a lot of audio. At 16 kHz mono PCM that's ~32 KB/second; at 22 kHz Opus that's ~10 KB/second. For a voice agent doing 30M characters/month (~1 million seconds of audio), egress is 10-30 GB/month, which most vendors include in pricing but some bill separately.
- Logging and audit retention: vendors bill differently for retention beyond the default. If you need 7-year retention for compliance (financial services, healthcare), the additional retention cost is substantial.
- Multi-region failover: vendors with regional deployments (UAE, Saudi, EU) sometimes charge a premium for failover across regions or for active-active deployments. Voice agents that can't tolerate single-region latency in a regional outage need this.
- Custom contractual SLAs: P0/P1 incident response below the public SLA tier, dedicated support engineer, custom uptime guarantees. Priced as a separate enterprise SKU.
- Data residency premiums: deploying the inference path in a specific jurisdiction (UAE-only, EU-only) sometimes carries a premium over the default global pool.
How to compare vendors
A like-for-like comparison requires normalising all five axes. The framework:
- Pick a representative workload (e.g. 10M characters/month batch + 1M characters/month streaming + 5 cloned voices).
- Get a written quote from each vendor for the workload at the volume you're projecting in 12 months.
- Add the hidden costs (egress, retention, custom SLAs, data residency) for each vendor.
- Normalise per-feature charges (dialect surcharges, SSML premiums, sample-rate premiums) into a single comparable price per-character.
- Compare the resulting fully-loaded cost-per-character.
Vendors will resist this because the simple pricing-page number flatters them. Customers who insist on the fully-loaded comparison get more accurate procurement decisions.
The Hakim pricing posture
For full transparency on Hakim's pricing:
- Per-character pricing for both batch and streaming, published on /pricing with the volume tiers visible.
- No dialect surcharges; all 15 dialects price at the same per-character rate.
- Voice cloning charged as a per-clone fee plus the standard per-character rate on cloned-voice output (no premium over catalogue voices).
- Free tier ships the same model as the paid tier, with a 10,000 character/month limit.
- Volume discounts kick in at 1M characters/month; enterprise contracts (annual commits, reserved capacity, custom SLAs) negotiated above 100M characters/month.
- All output rates and formats included in the per-character price; no charge for higher sample-rate output.
The pricing is designed to be predictable from a script. If your workload is a 400,000-character audiobook, you can compute the cost on /pricing without a sales call.
A short pricing-evaluation checklist
When evaluating an Arabic TTS vendor's pricing:
- Get the full per-character price for batch and streaming separately.
- Ask if there's a dialect surcharge.
- Ask the voice-cloning fee structure: per-clone, per-character premium, or both.
- Ask if the free-tier model is the same as the paid-tier model.
- Ask the volume-discount break points and the rates at 10M, 100M, 1B characters/month.
- Ask which line items are NOT in the per-character price (egress, retention, SLAs, residency).
- Get a written quote for a representative 12-month workload, fully loaded.
- Compare the fully-loaded cost-per-character across vendors, not the public pricing-page number.
If you want to walk through the math for a specific workload, the /pricing page has a calculator; for enterprise volumes, email hello@tryhakim.ai with the projected workload and we'll send a written quote.