Text to Speech API — Batch, 30+ Languages, Pay per Character

Give it one text or a thousand, get one voice-over file per text (MP3 or WAV), one dataset item per text: the audio URL, its duration, the character count, or a clear error. English runs on Kokoro (fast, cheap); French, Spanish, German, Japanese and 30+ other languages run on MiniMax Speech 2.8 with natural voices. Language is detected automatically. No API key, no subscription. From $0.05 per 1,000 characters, failures are free.

Built for automation and AI agents: every input alias is accepted (texts, text, prompts, prompt, sentences), nothing is required, and the default input runs on its own so you can see the output shape before sending your own texts.

What you get

One item per text, in your dataset:

{
  "index": 1,
  "text": "Bienvenue dans notre boutique. Votre commande a été expédiée.",
  "characters": 61,
  "outputUrl": "https://api.apify.com/v2/key-value-stores/abc123/records/speech-0001.mp3",
  "durationMs": 4310,
  "format": "mp3",
  "language": "French",
  "voice": "Calm_Woman",
  "model": "fal-ai/minimax/speech-2.8-turbo",
  "processingMs": 3800,
  "status": "ok",
  "error": null
}

A text that fails (empty, over 5,000 characters, model error) comes back with "status": "error" and the reason in error. It is never charged.

Use cases

• Voice product descriptions, notifications or order updates in the customer's language.
• Turn articles, scripts or lessons into audio versions in bulk.
• Generate voice-overs for short videos from a content calendar.
• Give an AI agent a single call that turns a list of lines into a list of audio files.

Input

Field	Type	Default	Notes
texts	array	two sample sentences	Up to 5,000 characters each. For a single text, use text.
text, prompts, prompt, sentences	aliases	—	Merged with texts.
language	auto or a language name	auto	Detected per text when auto. English uses the standard model; other languages always use premium.
voice	female, male, or a model voice id	female	Kokoro ids (af_heart, am_michael…) for English standard; MiniMax ids (Wise_Woman, Deep_Voice_Man, Friendly_Person…) for premium.
quality	standard, premium	standard	premium forces MiniMax for every text, with pauses <#1.5#> and tags like (laughs).
speed	0.5 to 2.0	1.0
outputFormat	mp3, wav	mp3	Standard (Kokoro) always returns WAV; premium returns MP3.
maxTexts	integer	200	Stop after this many texts.

Minimal call:

{ "texts": ["Your package has been delivered.", "Votre colis a été livré."] }

Languages

36 languages: English, French, Spanish, German, Italian, Portuguese, Dutch, Polish, Russian, Turkish, Arabic, Hindi, Japanese, Korean, Chinese (Mandarin), Cantonese, Indonesian, Malay, Vietnamese, Thai, Swedish, Danish, Norwegian, Finnish, Greek, Czech, Slovak, Romanian, Hungarian, Ukrainian, Bulgarian, Croatian, Slovenian, Catalan, Hebrew, Afrikaans. Each text is detected on its own when language is auto; Malay and Cantonese must be forced, because detection cannot tell them apart from Indonesian and Mandarin. A language detected outside this list is still voiced by the premium model, which then detects it itself. Close languages can be confused (Croatian is often detected as Slovenian): force language when you know it.

Pricing

Charged per started 100 characters of text actually voiced:

• Standard (English, Kokoro) — $0.005 per 100 characters, i.e. $0.05 per 1,000.
• Premium (any language, MiniMax) — $0.012 per 100 characters, i.e. $0.12 per 1,000.
• Failed texts: free.

Set a maximum cost per run in Apify and the Actor stops cleanly when it is reached; remaining texts are reported, not processed.

Via API, n8n, Make or MCP

Call it like any Apify Actor: POST https://api.apify.com/v2/actors/sherwood~text-to-speech-batch/run-sync-get-dataset-items with the JSON input above, or add it as a tool through the Apify MCP server. Results are plain dataset items, so get-dataset-items returns the URLs directly.

Limits

• 5,000 characters per text; split longer scripts into paragraphs (one item each).
• Generation is typically 3 to 8 seconds per text; the standard model can take longer on its first call of a run while it warms up. 4 texts are processed in parallel.
• Output files live in the run's key-value store and follow your account's data retention.