Text to Speech Generator for Synthetic Audio Datasets

By SynthEval editorial team · Updated 2026-11-18

SynthEval provides a high-quality, API-first text-to-speech generator designed specifically for startups building machine learning pipelines. We deliver labeled synthetic audio datasets at scale, eliminating the licensing overhead and legal complexities associated with human voice recordings. Our platform is engineered to fuel model training with consistent, watermark-marked speech that ensures clear provenance for your training data.

This page outlines how to integrate our JSON API into your data generation workflows. You will find concrete details on our 8 licensed voices, 23 supported languages, and straightforward pricing. Whether you are augmenting sparse datasets or stress-testing voice recognition systems, our infrastructure offers the reliability and transparency needed for production-grade ML development.

Text to Speech Generator for Synthetic Audio Datasets

Frequently asked questions

What is a text to speech generator for ML?

It is a cloud API that accepts text and returns audio files for training pipelines. By sending a POST request to the speech endpoint, you receive synthetic speech. This allows startups to generate labeled audio datasets at scale without licensing overhead, using standard HTTP requests compatible with OpenAI SDKs.

How do synthetic voices compare to real recordings?

Our voices are based on licensed recordings from LibriTTS-R and LibriVox public-domain audiobook readings. Every generated file carries an inaudible watermark marking it as synthetic speech. This ensures traceability for your ML models while providing high-quality, consistent audio data for training purposes without the variability of human recording sessions.

Which voice is best for narrative datasets?

Luna is well-suited for soft storytelling, while Aria offers a warm, clear narrator style. Sage provides a calm, instructional tone. You can select any of the eight presets via the voice parameter in your API request to match your specific dataset requirements.

Are the voices based on public domain data?

Yes. The voices are based on licensed recordings from LibriTTS-R (CC BY 4.0) and from LibriVox public-domain audiobook readings. This foundation allows you to generate synthetic speech datasets without complex licensing overhead, ensuring your training data is derived from legally clear sources for broad commercial and research use.

Can I use the audio files for commercial training?

Yes, the generated audio files are designed for use in commercial training pipelines. The input text is used only to produce the audio and is not used for training our models. You receive the audio directly in the response, allowing you to integrate it into your ML datasets immediately.

How does the watermark work in synthetic audio?

Every generated audio file carries an inaudible watermark that marks it as synthetic speech. This watermark is embedded directly into the audio data returned by the API. It allows you to identify synthetic samples in your datasets, which is useful for mixed-source training or verifying the origin of your data.

What happens if the API request fails?

If synthesis fails, the charge is returned in full. Failed requests are free, so you are not billed for unsuccessful audio generation. This ensures that your API balance is only reduced when you successfully receive an audio file, protecting your budget during dataset creation experiments or high-volume generation phases.

Is there a limit on characters per request?

Yes, the input field accepts up to 4096 characters per request. For longer texts, you should split the content into parts of this size and join the resulting audio files. This limit ensures consistent processing performance, and you can manage larger datasets by programmatically chunking your text before sending requests.

Do I need to pay for every character generated?

You pay $0.01 per 1,000 characters, including spaces and punctuation. The rate is $10 per 1 million characters. Failed requests are free, so you only pay for successful generations. This predictable pricing model allows you to budget accurately for your ML dataset creation based on the exact character count of your text.

Can I adjust the speed of the generated speech?

Yes, you can adjust the speed parameter between 0.25 and 4.0. This allows you to generate speech at different rates to suit your dataset requirements, such as faster narration for efficiency or slower speech for clarity. The API processes these requests alongside other parameters like voice and language settings.

What formats are supported for audio output?

The API supports mp3, wav, opus, flac, and pcm via the response_format field. You can choose the format that fits your ML pipeline’s storage or processing needs. These formats are returned directly in the API response, allowing you to store or process the audio files immediately after generation.

How fast is the text to speech generation?

Generation is roughly real time on long texts, about 25 characters per second. Short phrases take a few seconds. Requests are processed one after another, so under load a request may wait in a queue. This predictable speed helps you estimate processing times for large-scale dataset generation projects.

Is there a free trial for new accounts?

Yes, signing in with Google provides $0.50 of credit, valid for 7 days, covering about 50,000 characters. Accounts created with email and password start without trial credit. This allows you to test the API and generate initial audio samples before committing to a top-up via USDT on the TRON network.

How do I top up my API balance?

Top up using USDT on the TRON network (TRC-20) only. The minimum amount is $5, and credits are automatically added within minutes after the transfer. There are no cards, PayPal, or subscriptions. Your balance does not expire, allowing you to fund your account once and use it for ongoing dataset generation.

First request

curl

curl https://api.play98f.com/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

API-First Synthetic Speech Generation

Sound wave illustration for: API-First Synthetic Speech Generation

SynthEval provides a cloud-based text-to-speech API designed for automated dataset generation. The endpoint POST https://api.play98f.com/v1/audio/speech accepts JSON bodies identical to the OpenAI speech endpoint, allowing startups to integrate immediately by simply updating the base_url in existing OpenAI SDKs.

Audio files are delivered synchronously upon successful synthesis. The service supports five standard formats: mp3, wav, opus, flac, and pcm. This immediate delivery ensures that audio assets are available for downstream processing without waiting for asynchronous job completion.

For large-scale dataset creation, text exceeding 4,096 characters is automatically split into segments. The API processes these sequentially, returning concatenated audio files. This approach maintains character limits while handling longer narration needs typical in training data pipelines.

JSON API compatible with OpenAI SDKs

The API endpoint mirrors the structure of the OpenAI speech endpoint, enabling drop-in integration for teams already using OpenAI SDKs. By changing the base_url to https://api.play98f.com/v1, developers can leverage existing codebases for synthetic audio generation.

Requests require a Bearer token in the Authorization header. Each account holds one active key, which can be rotated instantly. This simplicity reduces integration overhead for engineering teams focusing on model training rather than infrastructure management.

  • Compatible with official OpenAI SDKs via base_url change
  • Standard JSON request body with model, input, voice, response_format, and speed fields
  • Supports tts-1 and tts-1-hd models for varying quality needs

Instant audio file delivery (mp3, wav, opus, flac, pcm)

Upon successful synthesis, the API returns the audio file directly in the response. Supported formats include mp3, wav, opus, flac, and pcm, catering to various training pipeline requirements. Failed requests incur no charge, ensuring cost efficiency during development.

Generation speed is roughly real time on long texts, processing about 25 characters of text per second of audio. Short phrases take a few seconds, while longer texts are processed in real-time. Requests are handled sequentially, so high concurrency may introduce slight queue delays.

  • Formats: mp3, wav, opus, flac, pcm
  • Audio returned directly in the HTTP response
  • Failed requests are free

Batch processing for large dataset volumes

For texts longer than 4,096 characters, the system automatically splits the input into manageable segments. Each segment is synthesized independently, and the resulting audio files are joined to produce a single continuous output. This ensures consistent quality across long-form dataset entries.

With a rate limit of 60 requests per minute per key, startups can scale their data generation efficiently. The queue-based processing handles load gracefully, prioritizing sequential delivery of audio assets for ML training pipelines.

  • Automatic text splitting at 4,096 characters
  • Sequential audio joining for long-form content
  • 60 requests per minute standard rate limit

Licensed Voices for Training Data

Sound wave illustration for: Licensed Voices for Training Data

SynthEval’s audio engine relies on eight voice presets derived from the LibriTTS-R dataset, licensed under CC BY 4.0. This foundation provides startups with legally clear, high-quality synthetic speech for training machine learning models without the overhead of individual voice licensing negotiations.

8 presets derived from LibriTTS-R (CC BY 4.0)

Each voice is a licensed recording from LibriVox public-domain audiobook readings, ensuring a clean copyright chain for commercial dataset generation. The API exposes these eight distinct presets, allowing developers to mix and match tones across thousands of generated audio files with confidence.

  • 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

Inaudible watermarking for synthetic detection

Every generated audio file contains an inaudible watermark that marks the speech as synthetic. This feature enables downstream pipelines to easily distinguish between human-recorded and AI-generated audio, which is critical for maintaining data integrity in large-scale training sets.

Clear narrator, upbeat, storytelling, and documentary tones

The eight presets cover a wide emotional and functional range, from the neutral clarity of Atlas to the deep documentary style of Orion. Developers can select specific voices to match the demographic diversity required for robust speech recognition model training.

Licensed Voices for Training Data

Language Support for Multilingual Models

Sound wave illustration for: Language Support for Multilingual Models

Multilingual models require diverse phonetic data. Our API supports 23 languages, from Arabic to Turkish, ensuring comprehensive coverage for global speech recognition training.

23 supported languages from English to Turkish

Support covers Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, and Turkish. Each voice can speak any of these languages, allowing flexible dataset construction without voice-specific limitations.

Automatic language detection via script

If the language field is omitted, the system guesses the language from the script. Latin-script text defaults to English. This simplifies initial testing but may require explicit codes for mixed scripts.

Explicit language code override for Latin script ambiguity

For non-English Latin-script text, set the language parameter explicitly. This prevents misidentification and ensures accurate phonetic generation for languages like Dutch or Swedish that share the Latin alphabet.

Unsupported languages include Vietnamese, Indonesian, Thai, Ukrainian, Romanian, Czech, and Bengali. Do not assume support for these without verification.

Scalable Infrastructure for Startups

Sound wave illustration for: Scalable Infrastructure for Startups

SynthEval is engineered to handle the variable throughput demands of ML training pipelines without requiring complex infrastructure scaling. The API enforces a standard rate limit of 60 requests per minute per API key, which scales to 10 requests per minute for new accounts until the first top-up. This predictable ceiling allows data engineers to design reliable batch processing scripts that respect limits without overwhelming the service.

Longer texts are automatically split into 4,096-character segments, with audio files joined seamlessly. Generation speed approximates real-time, processing roughly 25 characters per second, though requests queue during high load. This queue-based processing ensures stability, allowing startups to generate large datasets consistently without manual intervention or managing server-side concurrency.

60 requests per minute standard rate limit

The API key supports 60 requests per minute under normal operation. New accounts start with a lower limit of 10 requests per minute until the first top-up is processed. This structure prevents sudden traffic spikes from degrading service quality for other users in the shared infrastructure.

Queue processing for longer text segments

Requests are processed sequentially. Under load, a request may wait in a queue before synthesis begins. This ensures fair resource distribution and consistent audio quality, even during peak usage periods common in large-scale dataset generation.

No monthly plans or minimum spend requirements

Startups can scale costs directly with usage. There are no monthly subscriptions, no seat limits, and no minimum spend requirements. You only pay for the characters you generate, making it easy to budget for experimental projects or production pipelines alike.

Data Privacy and Usage Rights

Sound wave illustration for: Data Privacy and Usage Rights

SynthEval’s API is engineered for data integrity. When you send input text, it is processed strictly to produce the requested audio file and is not added to our training pools. This ensures your proprietary prompts remain isolated from model development cycles.

Once the synthesis completes, the audio response is delivered directly to your client. We do not store the generated audio files by default, minimizing your storage overhead and ensuring your synthetic datasets are truly ephemeral unless you choose to save them locally.

Input text used only for generation, not training

Your input text is consumed solely to generate the audio output for that specific request. It is not ingested into our voice models or used to improve the base engine. This separation guarantees that your unique dataset prompts do not influence future model behaviors.

Audio returned in response, not stored by default

The generated audio file is returned in the HTTP response for you to save. We do not maintain a permanent copy of the audio on our servers after delivery. This approach reduces our storage costs and keeps the ownership of the synthetic assets entirely with you.

Metadata kept solely for billing accuracy

We retain request metadata—such as timestamp, character count, and status—exclusively to verify your usage and calculate charges accurately. This data is not used for marketing or model training. It ensures transparent billing, reflecting the exact $10 per million characters rate.

  • Timestamps for audit trails
  • Character counts for billing
  • Request status for troubleshooting
Data Privacy and Usage Rights

Developer Experience

Sound wave illustration for: Developer Experience

Integration is straightforward: generate an API key via the dashboard, then send a POST request to the speech endpoint. The JSON body accepts standard fields, allowing you to swap in existing OpenAI SDKs by simply adjusting the base URL to our API domain.

Simple cURL examples for immediate testing

Start with a single cURL command to verify connectivity. Send your text payload along with the required voice and model parameters to receive an audio file directly in the response, enabling quick validation of your data generation pipeline.

Support for tts-1 and tts-1-hd models

Choose between two identical-engine models: "tts-1" for standard throughput or "tts-1-hd" for higher fidelity. Both accept the same input structure, allowing you to experiment with quality levels without changing your integration code.

Speed control from 0.25 to 4.0

Adjust the generation speed between 0.25x and 4.0x to match specific acoustic requirements. This parameter helps you create diverse speech rates for robustness testing across different voice profiles without needing external audio processing tools.

Use Cases for Synthetic Speech

Sound wave illustration for: Use Cases for Synthetic Speech

Synthetic speech APIs solve the data scarcity bottleneck in machine learning by providing labeled audio at scale. Unlike licensed recordings, generated datasets avoid complex rights management, allowing startups to train models on diverse, high-quality audio without licensing overhead.

This approach enables rapid iteration. Teams can generate thousands of variations instantly, ensuring their models generalize well across different accents, speeds, and vocal characteristics before moving to production.

Augmenting sparse language datasets

For languages with limited available data, synthetic generation fills the gap. By leveraging 23 supported languages, you can create balanced training sets for under-resourced dialects or low-resource environments where commercial datasets are expensive or unavailable.

You can generate specific phoneme combinations or edge-case pronunciations that rarely appear in existing corpora, improving model robustness without waiting for human annotators.

  • Generate specific phoneme combinations for under-resourced languages
  • Create balanced datasets for low-resource dialects
  • Reduce annotation costs by automating labeling

Stress testing voice recognition pipelines

Automated pipelines require robust handling of edge cases. Synthetic speech allows you to inject noise, vary speeds from 0.25x to 4.0x, and test recognition accuracy across all 8 voice presets systematically.

This ensures your ASR models handle variations in pitch and cadence without requiring manual recording sessions for every edge case.

Failed synthesis requests are free, allowing you to generate thousands of test samples without cost penalties if errors occur.

Generating diverse speaker profiles for bias testing

Bias in speech models often stems from homogeneous training data. Using 8 distinct voices (female and male presets) helps you audit your models for gender, tone, and style biases.

You can generate datasets with controlled variables to isolate how specific vocal traits affect recognition accuracy, ensuring fair performance across different speaker profiles.

The inaudible watermark helps identify synthetic data in training sets, allowing you to track the proportion of synthetic vs. real data in your models.

Pricing Structure for Data Engineers

Sound wave illustration for: Pricing Structure for Data Engineers

Our pricing is linear and predictable, designed for data engineers who need to forecast dataset costs. There are no monthly plans, seats, or minimum spend requirements, allowing you to scale usage up or down based on your training pipeline needs without administrative overhead.

The rate is $10 per 1 million characters, equating to $0.01 per 1,000 characters. Failed requests are free, ensuring you only pay for successful audio generation. A free trial of $0.50 credit is available for new accounts via Google sign-in.

Flat rate of $10 per 1 million characters

The cost scales directly with volume. At 60 requests per minute, you can generate significant data rapidly. Use the dashboard to monitor your balance and characters used in real-time.

  • No tiered pricing structures
  • Balance does not expire

Spaces and punctuation count toward limits

Character counting is strict and inclusive. Every space, comma, and period contributes to the 1 million character threshold. This ensures accurate billing regardless of text formatting or spacing variations in your input JSON.

Free credits for failed synthesis requests

If synthesis fails, the charge is returned in full. This protects your budget against transient API errors or invalid inputs, ensuring you only pay for delivered audio files.

Pricing Structure for Data Engineers

Integration Workflow

Sound wave illustration for: Integration Workflow

Integration begins with account creation via Google or email on the /get-api-key/ page. Once registered, you access the dashboard to view your unique API key and current balance.

The dashboard displays your key, usage statistics, and a ready-to-use cURL example, allowing immediate testing without additional configuration steps.

Accounts are created per site, ensuring isolated credentials for your synthetic dataset projects.

Sign up with Google or email password

Create your account using Google OAuth or a standard email and password combination. This process is performed on the /get-api-key/ page specific to play98f.com.

Google sign-ins receive a $0.50 trial credit valid for 7 days. Email-based accounts start without trial credit but can top up via USDT (TRC-20).

Your account is tied to this specific site, ensuring clear separation of credentials and billing history.

  • Google OAuth or Email/Password options
  • $0.50 trial credit for Google sign-ins
  • Instant account creation

Retrieve API key from dashboard

Upon login, the dashboard displays your active API key, starting with 'tts-'. You can issue a new key at any time, which immediately invalidates the previous one.

The interface shows your balance, characters remaining, and a summary of usage. This transparency helps you monitor dataset generation costs without needing external billing tools.

Keys are single-active-per-account, simplifying authentication management for CI/CD pipelines or automated dataset scripts.

  • Active key displayed prominently
  • Instant key rotation available
  • Balance and usage stats visible

Send POST request to /v1/audio/speech

Use the retrieved key in the Authorization header: 'Bearer tts-...'. Send a JSON payload to the endpoint with your text, selected voice, and desired format.

The API accepts standard OpenAI SDK parameters. Simply configure your SDK's base_url to https://api.play98f.com/v1 and use the standard speech endpoint.

Ensure your input text is under 4096 characters. For longer corpora, split your text into chunks, process them sequentially, and concatenate the resulting audio files.

Listen to the 8 voices

Each sample below was generated by this API from the same sentence. Our synthetic speech datasets help startups train smarter AI models without expensive voice talent.

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. Creative Commons Attribution 4.0
  3. Node.js documentation
  4. MDN: Web Speech API
  5. FFmpeg documentation
  6. LibriTTS-R paper (arXiv 2305.18802)
  7. openai-node on GitHub
  8. curl documentation
  9. TRON developer hub: TRC-20
  10. Tether (USDT)
  11. Python documentation
  12. OpenAI API reference: create speech

More on this site