Synthetic Speech API for ML Dataset Generation

By VoxyData editorial team · Updated 2026-11-19

VoxyData provides a reliable text-to-speech API designed to help product teams and ML engineers generate high-quality, copyright-clean synthetic audio datasets. By leveraging licensed recordings from LibriTTS-R and LibriVox, we ensure consistent voice quality and clear narration styles for your models.

This page explains how to use our transparent, character-based pricing to scale your training data efficiently. You will find concrete examples of technical integration, supported languages, and specific use cases for voice assistant training and TTS model fine-tuning.

Synthetic Speech API for ML Dataset Generation

Frequently asked questions

Is the synthetic speech suitable for training ML models?

Yes. VoxyData provides high-quality, copyright-clean synthetic audio datasets for voice models. Each generated file carries an inaudible watermark marking it as synthetic speech. Input text is used only to produce audio and is not used for training, ensuring clean data for your ML pipelines.

Do you use the input text for training your own models?

No. Input text is used exclusively to produce the audio response. It is not used for training our models. Generated audio is returned in the response and not stored. We only keep request metadata, such as time, characters, and status, for billing purposes.

Can I generate large datasets with consistent voice quality?

Yes. We provide consistent voice quality for dataset generation. You can split longer texts into parts of up to 4096 characters and join the resulting audio files. This allows you to build large, consistent datasets using our reliable TTS API with transparent character-based pricing.

Which voices are best for narrative datasets?

For narrative datasets, aria (warm, clear narrator), luna (soft, storytelling), or milo (friendly, conversational) are suitable. These voices are based on licensed recordings from LibriTTS-R (CC BY 4.0), from LibriVox public-domain audiobook readings, providing a natural foundation for synthetic speech datasets used in machine learning applications.

How do I handle text longer than 4096 characters?

Split your text into parts of up to 4096 characters. Send each part as a separate request to the speech endpoint. Then, join the resulting audio files. This workflow ensures you can generate audio for texts of any length while adhering to the character limit per request.

Are the voices based on public domain recordings?

Yes. Every voice is based on licensed recordings from LibriTTS-R (CC BY 4.0), from LibriVox public-domain audiobook readings. This ensures your synthetic datasets are built on clean, well-documented source material, suitable for commercial and research use without additional licensing concerns for the base recordings.

What audio formats are supported for data ingestion?

You can choose from mp3, wav, opus, flac, and pcm. These formats are available via the response_format parameter in your API request. Select the format that best fits your data ingestion pipeline, ensuring compatibility with your machine learning framework’s audio processing requirements.

Does the API support multiple languages in one request?

No. Each request processes a single language. You can optionally set the language using an ISO code. If not set, it is guessed from the script, defaulting to English for Latin text. You must send separate requests for different languages within a dataset.

How fast is the speech generation for bulk processing?

Generation is roughly real time on long texts, about 25 characters per second. Short phrases take a few seconds. Requests are processed sequentially, so under load, a request may wait in a queue. This predictable timing helps you plan your bulk processing workflows accurately.

Is there a limit on the number of requests per minute?

Yes. The rate limit is 60 requests per minute per key. Before your first top-up, this limit is 10 requests per minute. This ensures fair usage and consistent performance for your synthetic dataset generation tasks, allowing you to scale your production pipeline effectively.

Can I use the generated audio commercially?

A: Yes. The voices are based on licensed recordings from LibriTTS-R (CC BY 4.0), from LibriVox public-domain audiobook readings. VoxyData provides these as synthetic datasets for your models. The generated audio carries an inaudible watermark, clearly marking it as synthetic speech for transparency in your datasets.

What happens if a request fails? Do I get charged?

No. Failed requests are free. If synthesis fails, the charge is returned in full. You only pay for successful character processing. This ensures you are not billed for audio files that were not successfully generated, keeping your dataset generation costs accurate and predictable.

Do you offer custom voice cloning for datasets?

No. We do not offer custom voice cloning, voice design, or uploading your own voice. You choose from eight preset voices based on licensed recordings. This ensures consistent, copyright-clean synthetic audio for your ML datasets.

How do I track my character usage and billing?

Sign up at /get-api-key/ to view your balance, trial status, characters left, and usage history. Request metadata (time, characters, status) is kept for billing. You can also issue a new API key at any time, which immediately invalidates the old one for security.

First request

curl

curl https://api.vd89.info/v1/audio/speech \
  -H "Authorization: Bearer $TTS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"tts-1","voice":"aria","input":"Hello from the API.","response_format":"mp3"}' \
  --output hello.mp3

Why synthetic speech matters for ML

Sound wave illustration for: Why synthetic speech matters for ML

Synthetic speech enables product teams to generate large-scale, copyright-clean audio datasets efficiently. By automating voice generation, you eliminate the logistical bottlenecks of traditional recording sessions, allowing for rapid iteration of voice models without geographic or scheduling constraints.

Scale your training data

Scale your training data

Generate thousands of hours of audio on demand. With a rate limit of 60 requests per minute, you can systematically build diverse corpora by varying text inputs and parameters, ensuring your ML models are trained on robust, varied data.

Reduce collection costs

Avoid the overhead of studio rentals, talent fees, and post-production. At $0.01 per 1,000 characters, synthetic speech offers predictable, character-based pricing that scales linearly with your dataset size, significantly lowering the cost per hour of audio compared to human recording.

Control voice attributes

Select from eight distinct voice presets, each with a specific tonal character. Adjust speed between 0.25x and 4.0x to match specific training requirements. This level of control ensures consistent audio quality and narration style across your entire dataset.

Voice catalog for model training

Sound wave illustration for: Voice catalog for model training

VoxyData provides eight distinct voice presets optimized for synthetic dataset generation. Each voice offers a unique tonal character, allowing product teams to diversify their training corpora without the variability of human recording sessions.

Diverse voice presets

Our catalog includes four female and four male voices, ranging from warm narrators to energetic promoters. This variety ensures your models encounter a broad spectrum of vocal timbres and pacing, reducing bias in voice recognition systems.

  • Aria: Warm, clear narrator
  • Nova: Bright, upbeat
  • Luna: Soft, storytelling
  • Sage: Calm, instructional
  • Felix: Energetic, promo
  • Milo: Friendly, conversational
  • Atlas: Neutral, news
  • Orion: Deep, documentary

Consistent audio quality

Every generated file maintains uniform audio characteristics, critical for training robust ML models. The voices are based on licensed recordings from LibriTTS-R (CC BY 4.0), from LibriVox public-domain audiobook readings. Each file includes an inaudible watermark to identify synthetic speech during dataset curation.

Clear narration styles

The eight presets cover distinct narration styles, from neutral news delivery to deep documentary tones. This clarity helps models distinguish between speaking styles and emotional contexts, improving performance in voice assistants and TTS fine-tuning tasks.

Technical integration

Sound wave illustration for: Technical integration

Integrate VoxyData’s TTS engine directly into your data pipelines using our standard REST API. Send text via JSON and receive audio files, making it easy to automate large-scale dataset generation for voice models.

Simple REST API

Use a simple POST request to generate speech. The API accepts text up to 4096 characters and returns audio files directly, allowing seamless integration into existing ML workflows without complex middleware.

  • JSON-based request body
  • Standard HTTP methods

OpenAI SDK compatible

Leverage existing OpenAI SDKs by updating the base URL to https://api.vd89.info/v1. This compatibility lets you switch engines without rewriting your integration code, speeding up development for product teams.

  • Compatible with official OpenAI SDKs
  • Minimal configuration changes
  • Familiar endpoint structure

Standard audio formats

Choose from mp3, wav, opus, flac, or pcm based on your model’s requirements. High-quality outputs ensure your synthetic datasets maintain clarity, while flexible formats reduce post-processing needs.

  • mp3, wav, opus, flac, pcm
  • Configurable via API
  • Consistent bitrates

Language support

Sound wave illustration for: Language support

VoxyData supports 23 languages, including English, Chinese, Spanish, Arabic, Hindi, and Japanese. Each voice preset can speak any supported language, allowing you to generate diverse training data without managing multiple voice models.

23 supported languages

Our API covers a broad range of global languages, from European and Asian languages to African and Middle Eastern languages. You can generate high-quality synthetic speech for any of the 23 supported languages, ensuring your dataset reflects real-world linguistic diversity.

  • Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Turkish.

Multi-voice flexibility

Every voice preset is capable of speaking all 23 languages. This means you can maintain consistent voice characteristics across different languages or switch voices easily within a single project. There is no need to assign specific voices to specific languages.

Automatic language detection

If you do not specify a language, the API guesses it from the text script. Latin script defaults to English, but for non-English Latin text, you should set the language parameter explicitly to ensure accurate pronunciation.

Pricing for data generation

Sound wave illustration for: Pricing for data generation

Synthetic dataset generation requires predictable, volume-based costs. VoxyData charges $10 per 1 million characters, meaning every 1,000 characters cost just $0.01. Spaces and punctuation count toward this total, giving you precise control over your budget when scaling audio training data.

Pay per character

Billing is strictly character-based. Failed requests are free, so you only pay for successfully generated audio files. This ensures you are not charged for noise or errors during bulk dataset creation.

No minimum spend

There are no monthly plans, seats, or minimum spend requirements. You can top up with as little as $5 using USDT (TRC-20) on the TRON network. Balance does not expire, allowing you to generate datasets at your own pace.

Transparent billing

Request metadata, including time, character count, and status, is kept for clear billing. You can track usage directly in your account dashboard. No hidden fees or VAT invoices are involved; just straightforward, usage-based billing.

Pricing for data generation

Dataset quality features

Sound wave illustration for: Dataset quality features

Synthetic speech datasets require distinct quality markers to ensure models distinguish generated audio from real recordings. VoxyData embeds an inaudible watermark in every generated file, clearly marking it as synthetic speech without affecting perceptual quality or model training performance.

Inaudible watermarks

All audio outputs carry a digital watermark that identifies them as synthetic. This allows your pipeline to filter or tag training data accurately, ensuring your models learn from the correct source type without manual labeling overhead.

Clean audio output

Generated files are returned directly in the response, ensuring clean audio output without storage artifacts. The API delivers standard formats like mp3 or wav, ready for immediate ingestion into your training pipelines without additional preprocessing.

No training data usage

Input text is used solely to produce the audio and is not stored for training purposes. Generated audio is returned in the response and not retained, ensuring your input data remains private and your training datasets remain distinct from your API usage logs.

Use cases for product teams

Sound wave illustration for: Use cases for product teams

Synthetic speech enables product teams to build robust voice interfaces without the logistical burden of recording hundreds of hours of human dialogue. By generating diverse utterances at scale, you can train models on a wider variety of phrasings, accents, and sentence structures than manual recording typically allows.

This approach ensures your voice assistants handle edge cases more gracefully. With consistent audio quality, teams can rapidly iterate on prompt engineering and model architecture before committing to expensive human data collection campaigns.

Voice assistant training

Generate thousands of unique query variations to cover long-tail intents and regional dialects. The API’s support for 23 languages allows you to train multilingual assistants without managing separate recording sessions for each market.

Use different voice presets to test how users respond to varying tones, from the calm instructional style of Sage to the energetic promo style of Felix, optimizing for engagement metrics.

  • Cover diverse intonations and sentence structures
  • Test multilingual support across 23 languages
  • Optimize voice tone for user engagement

TTS model fine-tuning

Create clean, watermark-labeled training sets to fine-tune custom models without worrying about copyright clearance. Since input text isn’t used for training and audio isn’t stored, you maintain full control over your dataset’s provenance.

Leverage the consistent engine quality of tts-1 and tts-1-hd to generate uniform datasets. The inaudible watermark helps distinguish synthetic data from real recordings during evaluation phases.

  • Copyright-clean synthetic datasets
  • Inaudible watermark for data provenance
  • Consistent engine quality for uniform training data

Audio content generation

Produce large volumes of narration for e-learning modules, IVR systems, or interactive stories. With speeds adjustable from 0.25x to 4.0x, you can create variations for different pacing requirements without re-recording.

Use standard formats like MP3 or WAV to integrate directly into existing media pipelines. The simple request structure keeps the generation process fast and predictable for automated content workflows.

  • Adjustable speed for different pacing needs
  • Standard audio formats for easy integration
  • Automated content workflows for e-learning and IVR

Getting started quickly

Sound wave illustration for: Getting started quickly

Create your API key in seconds via the /get-api-key/ dashboard. You receive a Bearer token immediately, with a $0.50 trial credit if you sign in with Google, valid for 7 days.

API key setup

Authenticate requests using the Authorization: Bearer header. If you need to rotate your key, issuing a new one instantly invalidates the previous one, ensuring secure access for your dataset pipelines.

First request example

Send a POST request to https://api.vd89.info/v1/audio/speech with your text and voice preset. The API returns the audio file directly, ready for integration into your training set.

Use the official OpenAI SDK by changing the base_url to https://api.vd89.info/v1. This allows you to leverage existing client libraries without rewriting your integration logic.

  • Endpoint: POST /v1/audio/speech
  • Auth: Bearer tts-...
  • SDK: Compatible with OpenAI Python/JS clients

Developer documentation

Full endpoint details, parameter descriptions, and rate limit specifics are available in the /docs/ section.

Explore the voice catalog and language support to ensure your synthetic data matches your model requirements.

Frequently asked questions

Sound wave illustration for: Frequently asked questions

Synthetic speech offers consistent, clean audio ideal for training voice models without the noise and variability of human recordings.

Our API generates standardized files with inaudible watermarks, ensuring data provenance while keeping input text off training sets.

Dataset suitability

Generated audio is returned in the response and not stored, while input text is used only for synthesis. Request metadata is retained for billing. Files carry inaudible watermarks marking them as synthetic speech.

  • Input text is not used for training.
  • Generated audio is not stored after response.
  • Inaudible watermarks identify synthetic speech.

Voice variety

We provide eight presets (e.g., Aria, Felix, Atlas) spanning warm, bright, soft, and neutral tones. Any voice can speak all 23 supported languages, allowing flexible dataset construction without voice cloning.

For non-English Latin text, set the language parameter explicitly. Otherwise, the system guesses from the script, defaulting to English.

  • 8 voice presets available.
  • All voices support 23 languages.
  • No custom voice cloning offered.

Technical limits

Each request accepts up to 4096 characters. Split longer texts into parts and join the resulting audio files. Rate limits are 60 requests per minute per key.

Speed is roughly real-time (25 characters per second). Short phrases may take a few seconds due to queue processing.

  • Max 4096 characters per request.
  • 60 requests per minute per key.
  • Real-time speed; queue processing applies.

Listen to the 8 voices

Each sample below was generated by this API from the same sentence. This helps prevent overfitting and improves model generalization.

Avatar of the voice aria
ariafemale
Avatar of the voice nova
novafemale
Avatar of the voice luna
lunafemale
Avatar of the voice sage
sagefemale
Avatar of the voice felix
felixmale
Avatar of the voice milo
milomale
Avatar of the voice atlas
atlasmale
Avatar of the voice orion
orionmale

23 languages

Any voice speaks any of these languages.

Sources and further reading

  1. Wikipedia
  2. LibriVox public-domain audiobooks
  3. openai-node on GitHub
  4. OpenAI API reference: create speech
  5. Node.js documentation
  6. FFmpeg documentation
  7. LibriTTS-R paper (arXiv 2305.18802)
  8. Creative Commons Attribution 4.0
  9. Python documentation
  10. IETF RFC 7845: Ogg Opus
  11. curl documentation
  12. W3C WAI: making audio and video accessible

More on this site