CastReader Voice APITTS API SELECTION · PRIMARY SOURCES
Six providers. Different meters, limits and workflows. Start with your requirements, compare the actual text cost, then verify the sound.
Compiled by CastReader. Competitor facts checked： · Official documentation comparison, not an independent audio benchmark.
Prepared learning clips, product guidance and short voiceovers with authorized cloned voices. Current deployment: 10 languages, 500 characters/job, concurrency 1.
If streaming conversation, native multi-speaker orchestration, emotion controls or a contracted latency SLA are mandatory, do not select this CastReader offering on price alone. We do not currently provide those capabilities.
Prices are USD list rates, not a binding quote. Trials, taxes, discounts, top-up minimums and account terms may change the amount you pay. Unknown means verify—not unsupported.
Short, prepared multilingual audio and authorized voice cloning. Saved MP3/WAV; recoverable queued jobs.
Access: self-service; 3,000 trial characters. Verify identity; one grant per account, validity and daily limits apply.
Current configuration: en, zh, de, ja, fr, es, ko, pt, ru, it; 500 characters/queued job, concurrency 1. No streaming, realtime, emotion parameter or latency SLA.
A candidate when faster delivery, more languages or longer individual inputs are required.
Pay-as-you-go is available; check free/startup allowances and account terms separately.
Official Flash/Turbo listing: 32 languages, 40,000-character input limit. v2/v3 have different prices and capabilities; not interchangeable benchmarks.
Official Python/TypeScript clients, cloning, REST and WebSocket streaming.
Pay-as-you-go. The official price list also includes s2.1-pro-free at $0; verify free-model access and limits before choosing paid usage.
Bytes are not characters. The pricing table lists starter concurrency 5; account/model-specific free limits and production terms require confirmation.
Consider alongside other speech-service options; evaluate the chosen voice and language in your own workload.
PAYG and Growth tiers have different rates. Free credits and other models are separate.
Aura-1 is also listed at $15/1M characters. Flux has a separate price/promotion; this row does not mix these models or establish clone parity.
Evaluate when its voice/cloning workflow matches the application; compare plan utilization, not only the headline fee.
A Free plan exists. Pro includes instant voice cloning and commercial use in the published plan.
No character-equivalent estimate here: confirm model credit conversion, overage, term and account limits first.
Streaming speech and delivery instructions; a candidate when an existing application already uses OpenAI.
API access and rate limits depend on the account. Custom voices have a separate eligibility and consent process.
2,000 input tokens maximum on the model page. Output audio tokens are not known from character count alone; no fixed $/1M-character conversion.
SAME TEXT · DIFFERENT METERS
Runs in your browser. No text is uploaded and no audio is generated.
1–1,000,000 whole requests. This models repeating the text, not one oversized request.
All estimates use the same explicitly preprocessed text: CRLF → LF, NFC normalization, outer whitespace trimmed. Internal spaces and punctuation remain.
Paid-rate illustration, not “lowest price.” Fish also lists s2.1-pro-free at $0; check its access and limits. Trial credits, subscriptions, discounts, taxes and minimum top-ups are not included.
ElevenLabs/Deepgram estimates assume one billed character per Unicode code point; verify their exact metering. Credits and output audio tokens require their own usage data. This is not equal-quality or equal-speed testing.
Install the exact published client, check credentials and the ready voice, then save the request and idempotency key before submitting. A timeout is not permission to create another billed job.
Run the connection check →Our expanded first-party run records 30 attempts, 29 saved outputs and one failure. It is not a competing-model benchmark or a native-speaker verdict. Inspect the original reference, generated audio, input and usage.
Inspect the actual report →total_cost = successful_audio_usage
+ deliberate_regenerations
+ subscription_or_unused_minimums
+ app_hosting_storage_and_delivery
+ integration_and_recovery_work
// Measure complete-job time separately from first-audio time.
// Do not infer output audio tokens from text character count.