Narakeet

Narakeet turns text, Markdown and PowerPoint-style presentations into narrated audio and video. Its strongest buying distinction is not the advertised 900 voices. It is an unusually explicit production contract: top-up credits do not expire, output is metered by the final second, commercial previews are free, and API length/rate/batch limits are public.

> Distinctive strength: Final-second billing, non-expiring top-up credits, free commercial previews and published API limits make production economics unusually testable. > > Where it stops being an advantage: Transparent billing does not prove that every advertised voice is equally natural or controllable.

BenPicks has not listened to a representative Narakeet cohort or measured voice quality. We do not claim that 900 voices are equally natural or controllable. This review shows buyers how to avoid billable rebuilds, test the exact voice/provider and calculate accepted-minute cost.

The Narakeet decision in 60 seconds

How much does Narakeet cost?

Narakeet's public top-up plans are one-time purchases, not recurring subscriptions. The pricing documentation identifies a $6 package for 30 minutes ($0.20/minute), packages through 10,000 minutes and a public range down to roughly $0.05/minute at volume. Confirm the live package table and currency; USD, GBP, EUR and INR are supported.

For text-to-speech and generated video, credits are deducted from final output duration in one-second increments. For transcription, source-file duration is charged. Credits do not expire.

The decisive rule: every full Create Audio/Create Video build is charged whether or not it is downloaded. If a 20-minute narration has the wrong pronunciation and you rebuild it twice, the account consumes 60 minutes. Use the free Preview function available to commercial accounts before each billable build.

Calculate:

`top-up cost ÷ (generated minutes − rejected/rebuilt minutes) + editor/listening cost per accepted minute`

Free accounts allow 20 conversions but lack commercial use, API, SSML, batch creation and free preview. They are useful for interface evaluation, not for producing publishable assets.

What limits change the workload?

The free/commercial comparison is concrete:

Narakeet says commercial limits can be increased on request. Treat that as a negotiation path, not an entitlement until written into the account.

Are commercial rights included?

Yes—for content generated after a commercial account is activated. Narakeet says paid Top-Up and Subscription accounts may use generated material online, offline, in broadcasts, radio and remixes without continuing voice royalties. Free-account material is personal/evaluation only.

That license covers the supplied voices, not your source text, slides, images, video, music, brands or third-party excerpts. You must own or license those. Platform monetization rules also remain separate.

This timing matters: buying a plan later does not necessarily cure output created under the free account. Create the final material while the account is commercial and retain the invoice, source-rights file, script and build record.

What do “900 voices in 100 languages” actually mean?

Narakeet aggregates voices from AWS Polly, Google, IBM, Yandex, Microsoft, CereProc, Unreal Speech, Rime and Inclusive Solutions. The catalogue therefore contains multiple engines with different behavior.

Controls include voice, speed, volume, pauses, emphasis, multiple speakers/languages and time-based synchronization. Pitch can be set from -150 to 150, but Narakeet says about 80% of voices support it and individual ranges vary. SSML and phonetic styles such as IPA, X-SAMPA, Pinyin and Jyutping are also per-voice; some support none.

Do not choose from a language count. Build a short test containing brand names, acronyms, dates, currency, addresses, emotional punctuation and required phonemes in each target voice. Record exactly which controls work.

How should free previews be used?

Preview a representative segment before the full build. Duration estimates derive from words and selected voice but can drift when scripts contain pauses, timing directions, custom speeds or pronunciation instructions. An exact preview of a representative section gives a better forecast.

Fix pronunciation with literal/IPA where supported, tune speed, then preview again. Only create the full asset after a second reviewer approves the sample. This workflow has direct economic value because preview does not deduct credits while every full rebuild does.

What are the API limits?

API keys are available to Top-Up and metered subscription commercial accounts. Free and unmetered flat-price subscriptions cannot use the API.

Narakeet exposes two text-to-speech paths:

• Streaming: up to 1 KB, intended for short content around 30 seconds, returns MP3/M4A bytes directly.
• Polling: up to 1024 KB and up to 45-minute builds, returns a status URL and also supports uncompressed 16-bit PCM WAV.

The default account quota is 86,400 requests/day—one per second. Batch ZIP generation supports 500 scenes/files by default. Narakeet explicitly recommends batching rather than firing many small parallel calls that can hit rate limits.

The Markdown-to-video API can package scripts and assets, but direct PowerPoint automation is not supported. This distinction matters for teams whose source of truth is a presentation deck rather than Markdown.

Test retries, duplicate jobs, polling expiry, invalid content, quota exhaustion and download failures. A generated-file pipeline should retain its own job manifest rather than relying on temporary links.

Does subtitle-to-speech preserve timing?

The API accepts SRT and VTT, but aligns entire sentences and does not automatically compress speech when the voice is slower than the caption window. You may need to increase voice speed. That can harm intelligibility.

For each subtitle scene, measure timing overrun, required speed and native-listener comprehension. If speed exceeds your acceptable band, rewrite the target text or change voice rather than forcing audio into the slot.

What happens to uploaded data?

Narakeet's security page says files are stored in AWS us-east-1, transferred via HTTPS and stored with S3 server-side encryption. Other users cannot access them; internal support access is described as occurring only on explicit request.

Source assets are automatically deleted after 30 days. Generated audio/video/subtitle results are retained up to 48 hours. Earlier removal can be requested. Download links are temporary, so production systems must copy accepted output to owned storage.

Account information lasts for the account and can be removed at expiry/request, but fraud/security IP/email records may be retained indefinitely. The public material did not establish a complete training-use contract across Narakeet and every voice provider, current SOC/ISO evidence, SSO/SCIM/audit controls or uptime/support SLA. Obtain written answers for sensitive or critical workloads.

Who should shortlist Narakeet?

Shortlist it for documentation videos, course narration, multilingual scripted audio and CI-built media where repeatability and controllable costs matter. It is particularly attractive when teams can exploit free previews and keep script/assets as code.

Look elsewhere for low-latency dialogue, a self-serve cloned brand voice, strict regional hosting or a workflow requiring published enterprise assurance/SLA without negotiation.

Browse AI Voice software, read the commercial-use voice guide, and compare adjacent products through AI voice comparisons.

A fair Narakeet evaluation protocol

1. Use one rights-cleared script and one presentation containing names, numbers, acronyms, mixed languages and timing constraints.
2. Select three voices per language from different provider families when identifiable.
3. Preview difficult passages; record unsupported pitch, SSML and phonetic controls per voice.
4. Have native listeners blind-grade intelligibility, pronunciation, pacing, artifacts and suitability.
5. Build MP3, M4A and polling WAV; compare duration, loudness and downstream compatibility.
6. Convert SRT/VTT and measure timing overrun before changing speed or text.
7. Build a video from Markdown and presentation; log layout/synchronization corrections.
8. Rebuild one corrected asset and reconcile billed seconds exactly.
9. Exercise streaming, polling and 500-scene batch behavior plus duplicate/retry/quota failure.
10. Confirm output was created under a commercial account and retain source rights.
11. Verify output download, 48-hour expiry, source deletion and an early-removal request.
12. Obtain training, processors, assurance, support and SLA terms if risk requires them.
13. Calculate accepted-minute cost including rebuilds and reviewer time.
14. Require a second reviewer to reproduce the go/no-go decision from retained evidence.

Pass only when the selected voice/control combination meets written listening thresholds, subtitle timing remains usable, automation handles failure safely and accepted-minute cost beats the current production path. Catalogue breadth is not the result; an approved asset produced predictably is.