Low-Cost AI Text-to-Speech API: A Pricing Decision Guide

Short answer

The cheapest posted unit is not enough. Teams should compare the billable unit, quality tier, usable audio minutes, latency, concurrency, voice cloning, and support path before calling a text-to-speech API low cost.

Unreal Speech is a useful low-cost reference because its official pricing makes the tradeoff visible. Basic advertises $4.99 per month for the first six months with 3 million characters, against a $49 list price, and the vendor publishes $16 per million characters for Basic usage above the allowance. That can be unusually inexpensive for predictable batch narration. It is not a universal default: Unreal does not currently offer voice cloning, and its public pricing does not provide a plan-by-plan concurrency or support matrix.

The right comparison is therefore not cheapest vendor, but cheapest route that passes the workload, quality, capacity, rights, and support gates. A temporary free model, promotional bundle, or low overage rate is useful only if it survives the steady-state budget and production test.

Normalize the bill before comparing

Start with a representative monthly corpus rather than a round number. Include every production language, names and numbers, SSML, repeated generations during editing, failed requests, retries, and any text that is synthesized twice for caching or fallback. Keep the source text and expected audio duration so character-, byte-, credit-, and minute-based offers can be compared against the same work.

The billing unit changes the answer. Unreal Speech, MiniMax, ElevenLabs, and Speechify publish character-based routes, but their counting and packaging differ. Speechify says whitespace and SSML tags are excluded from billable characters. Fish Audio charges its paid models by UTF-8 bytes, so one visible character may consume one byte in basic English text or multiple bytes in other scripts. Cartesia packages credits and publishes approximate included TTS minutes. None of those units should be copied into one spreadsheet column without conversion.

Use a steady-state formula: base subscription plus expected overage plus voice-clone or voice-design fees plus any required support or enterprise contract. Then divide by accepted output minutes, not submitted text. Accepted minutes exclude unusable takes and include the cost of regenerations. Keep the introductory and free cases as separate scenarios so a temporary offer does not become the assumed year-two budget.

Quality and speed also change unit economics. MiniMax charges different rates for Turbo and HD. ElevenLabs prices Flash or Turbo below its Multilingual v2 or v3 route and describes Flash as the lower-latency option. A faster or cheaper tier is not a saving when it fails the pronunciation, expression, stability, or language test and forces more regeneration.

Decision table

What each route actually buys

Unreal Speech should be modeled in three phases: free evaluation, the six-month Basic promotion, and the post-promotion or overage state. The promotional package is the strongest reason to trial it for bulk English-heavy narration, accessibility audio, or other repeatable stock-voice work. The main procurement question is not the character rate; it is whether the available voices, language coverage, endpoint behavior, capacity, and support are sufficient without cloning.

SpeechifyAI Build is easier to place in a steady-state spreadsheet. The Starter package and overage share the same $10 per million character rate, while larger plans lower the overage rate and add support. Its API limits document gives paid Build audio 15 simultaneous requests and 20 sustained requests per second. That combination makes it a practical low-cost production candidate when transparent billing and a documented paid ceiling matter.

Fish Audio demonstrates why a zero price is not automatically the lowest production cost. The free S2.1 Pro route is explicitly temporary, fair-use, best-effort, and without an SLA; requests may also be used for model improvement. The paid route is a cleaner budget baseline at $15 per million UTF-8 bytes. For multilingual products, run the real corpus through a UTF-8 byte counter rather than assuming one byte per character.

Cartesia is the cleanest example of minute and capacity economics. Its plan table converts credits into approximate TTS minutes and pairs each tier with a concurrent-request limit. A team building interactive speech can therefore choose from measured peak active contexts and 429 behavior instead of monthly volume alone. The cheapest plan that covers total minutes may still be too small for a bursty agent workload.

MiniMax makes quality-tier pricing explicit. Turbo costs less than HD, while rapid cloning and voice design have separate per-voice charges. The published RPM and TPM limits help with request planning but do not answer how many long generations can run simultaneously. Treat the pay-as-you-go rate and an audio subscription as separate buying routes, then confirm account limits in the console before forecasting throughput.

ElevenLabs is not the lowest posted character rate in this group, but it exposes a useful speed-versus-quality boundary. Flash or Turbo costs half the Multilingual v2 or v3 character rate, and Flash has higher plan concurrency. That makes the decision measurable: use Flash for low-latency speech if it passes listening tests, or pay for the higher-quality route when pronunciation, expression, or long-form consistency justifies it.

Run a production-shaped cost test

1. Build one corpus with short conversational turns, long narration, numbers, names, abbreviations, punctuation, and every production language. Use the same approved stock or cloned voice category across providers.
2. Generate each item at least twice. Record submitted characters, billable characters or UTF-8 bytes, credits, accepted audio seconds, failed calls, free regenerations, and paid retries. Cache identical successful outputs so integration mistakes do not become recurring synthesis cost.
3. Measure time to first playable audio and full completion from the intended deployment region. Capture p50 and p95 latency, 429 responses, reconnects, and sustained throughput at the expected burst pattern. Do not substitute a vendor model-latency claim for end-to-end application latency.
4. Review pronunciation, speaker consistency, expressiveness, artifacts, and long-form stability without showing reviewers the vendor name. Count how often each route needs manual correction or regeneration; that quality tax belongs in the cost model.
5. Finish with cloning consent, commercial-use rights, data retention, deletion, DPA, SLA, incident support, and custom-capacity checks. A low-cost API that needs an unbudgeted enterprise agreement is not actually the low-cost route.

Final selection boundary

Keep Unreal Speech as the low-cost reference when the job is predictable high-volume synthesis with stock voices and the promotional, list-price, and overage scenarios all work. Move away from that reference when voice cloning, published concurrency, contracted latency, or a defined support tier is mandatory.

Use SpeechifyAI Build or paid Fish Audio when a transparent low paid rate is the main constraint, while remembering that their character and byte units are different. Use Cartesia when real-time concurrency and streaming behavior drive the bill. Use MiniMax when Turbo versus HD and per-voice creation costs define the decision. Use ElevenLabs when its model, voice, cloning, governance, or enterprise path earns the higher unit price. The final choice should be the least expensive route that passes the same corpus, load test, rights review, and steady-state budget.