SynthEval TTS API for ML Dataset Generation

By SynthEval editorial team · Updated 2026-11-18

This guide evaluates the SynthEval API for teams building ML training datasets. We assess voice variety, licensing clarity, and integration speed to help you choose a reliable synthetic speech provider. You will find concrete metrics on character pricing, rate limits, and format support tailored for batch generation workflows.

Designed for startups and data engineers, this page details how to generate labeled audio at scale without licensing overhead. We cover technical integration with OpenAI SDKs, language coverage across 23 tongues, and privacy policies. Read on to understand how watermarking and transparent pricing support your specific dataset requirements.

SynthEval TTS API for ML Dataset Generation

Frequently asked questions

Which TTS API is best for ML datasets?

Q: Which TTS API is suitable for ML datasets? A: SynthEval provides high-quality, watermark-marked synthetic speech ideal for training pipelines. Our API generates labeled audio datasets at scale without licensing overhead. With 23 languages and 8 distinct voices, it supports diverse model training needs efficiently.

How do SynthEval voices compare to others?

Our eight voices are based on licensed recordings from LibriTTS-R (CC BY 4.0) and LibriVox public-domain audiobooks. Each generated file carries an inaudible watermark marking it as synthetic speech, ensuring clear provenance for your ML datasets.

Are the voices licensed for commercial use?

Yes. The underlying voice recordings are based on LibriTTS-R (CC BY 4.0) and LibriVox public-domain readings. This allows you to generate commercial audio datasets without additional licensing overhead, provided you comply with the source licenses.

Does SynthEval support SSML?

No. SynthEval does not support SSML, phoneme dictionaries, or custom pronunciation lexicons. You send plain text up to 4096 characters. The API guesses the language from the script or uses the optional language parameter.

Can I clone voices with SynthEval?

No. We do not offer custom voice cloning, voice uploading, or voice design. You choose from eight preset voices: Aria, Nova, Luna, Sage, Felix, Milo, Atlas, and Orion. These are fixed presets based on licensed recordings.

What is the best voice for narration?

Aria (warm, clear narrator) and Luna (soft, storytelling) are excellent choices for narration. Sage (calm, instructional) also suits educational content. These presets deliver clear, natural-sounding speech ideal for generating labeled audio datasets. Since voice suitability depends on your specific project tone, we recommend testing samples to determine the best fit for your machine learning pipeline.

How many languages does SynthEval support?

We support 23 languages: Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, and Turkish. Any voice can speak any of these languages.

Is SynthEval cheaper than competitors?

SynthEval charges $0.01 per 1,000 characters ($10 per million), with no subscriptions, seats, or minimum spend. Failed requests are free. While we offer transparent, usage-based billing, we do not claim to be the cheapest option overall. Competitors may offer lower per-character rates but often include hidden fees or tiered structures. Our model ensures you pay only for what you use, with no monthly commitments.

Do I need a subscription for SynthEval?

No. There are no monthly plans, seats, or subscriptions. You pay only for what you use. Top-ups are done via USDT on TRON. Balance does not expire, and you can issue a new API key anytime.

How fast is SynthEval generation?

Typical speed is roughly real time, about 25 characters of text per second of generation. Short phrases take a few seconds. Requests are processed sequentially, so under load, a request may wait in a queue.

Can I use SynthEval for Siri-like voice data?

Yes. You can generate synthetic speech for voice assistant training. Our voices like Aria or Nova provide clear, natural speech. Each file includes an inaudible watermark to identify it as synthetic, which is useful for dataset labeling.

Does SynthEval store my input text?

No. Input text is used only to produce the audio and is not used for training. Generated audio is returned in the response and not stored. We keep request metadata (time, characters, status) only for billing.

What is the character limit per request?

The limit is 4096 characters per request. For longer texts, split them into parts of up to 4096 characters and join the resulting audio files. Spaces and punctuation count toward the character limit and cost.

How do I start using SynthEval?

Sign up at /get-api-key/ using Google or email. You receive a free trial with $0.50 credit. Use the API key with the OpenAI SDK or HTTP POST to https://api.play98f.com/v1/audio/speech to generate your first audio file.

Criteria for Evaluation

Sound wave illustration for: Criteria for Evaluation

Selecting a TTS API for ML dataset generation requires balancing voice diversity with licensing certainty. Synthetic audio must sound natural while remaining legally clear for commercial training pipelines. We evaluate how well a service provides varied, high-quality presets without hidden usage restrictions.

Integration speed and API reliability directly impact dataset production efficiency. Developers need predictable response times and straightforward error handling to scale bulk generation. We assess how easily the API integrates with existing Python workflows and how it handles concurrent requests under load.

Voice variety and quality presets

Eight distinct voice presets cover a wide range of tonal qualities, from warm narrators to energetic promo speakers. This variety allows ML teams to simulate diverse speaker profiles without needing dozens of unique voice models. Each voice is optimized for clarity, ensuring consistent audio quality across different training scenarios.

  • Eight distinct voice personalities (e.g., aria, nova, felix)
  • Optimized for clarity and natural intonation
  • Suitable for diverse speaker profile simulation

Licensing clarity for commercial use

Voices are sourced from LibriTTS-R under CC BY 4.0, providing clear commercial usage rights. This transparency reduces legal risk when integrating synthetic speech into proprietary ML models. The inaudible watermark further identifies generated audio as synthetic, aiding in dataset auditing and compliance verification.

  • Voices sourced from LibriTTS-R (CC BY 4.0)
  • Clear commercial usage rights
  • Inaudible watermarking for synthetic identification

API ease of integration and speed

The API follows the OpenAI speech endpoint structure, allowing immediate use with existing SDKs. Changing the base URL is sufficient for integration, reducing development time. Requests are processed sequentially, with typical speeds mirroring real-time generation for long texts, ensuring predictable throughput for bulk dataset creation.

  • OpenAI-compatible endpoint structure
  • Minimal code changes for integration
  • Predictable throughput for bulk generation

SynthEval: Licensed Voices and Transparency

Sound wave illustration for: SynthEval: Licensed Voices and Transparency

SynthEval provides high-quality synthetic speech by leveraging voices from the LibriTTS-R dataset, licensed under CC BY 4.0. This ensures that generated audio for ML training pipelines carries clear commercial usage rights, removing licensing ambiguity for startups scaling their datasets.

Voices sourced from LibriTTS-R (CC BY 4.0)

Our eight voice presets—ranging from warm narrators to energetic promos—are derived from licensed LibriTTS-R recordings. This foundation guarantees consistent, high-fidelity audio suitable for diverse machine learning applications without the overhead of negotiating individual voice licenses.

Inaudible watermarking for synthetic identification

Every generated audio file contains an inaudible watermark marking it as synthetic speech. This feature allows downstream ML models to easily distinguish between human-recorded and AI-generated data, improving training data purity and model robustness.

Clear character-based pricing model

Pricing is straightforward: $10 per 1 million characters, with spaces and punctuation counted. Failed requests are free, ensuring you only pay for successful synthesis. There are no monthly subscriptions or hidden fees, making cost prediction for large-scale dataset generation simple.

SynthEval: Licensed Voices and Transparency

Comparison with Other Solutions

Sound wave illustration for: Comparison with Other Solutions

SynthEval offers eight distinct voice presets rather than custom cloning. This approach provides consistent, high-quality synthetic speech for large-scale dataset generation without the variable quality of clone models. It is ideal for startups needing uniform audio characteristics across thousands of training samples.

Unlike SSML generators, we do not support phoneme dictionaries or emotion tags. Our API relies on straightforward text input and speed adjustments. This simplicity reduces parsing complexity for ML pipelines while maintaining high fidelity through licensed source material.

Difference from voice cloning services

SynthEval uses eight licensed presets derived from LibriTTS-R rather than offering custom voice cloning. This ensures consistent audio quality across all generated samples, which is critical for training robust machine learning models. Unlike clone services that may vary in fidelity per request, our presets provide reliable, standardized outputs for bulk dataset creation.

We do not allow customers to upload their own voice recordings. This limitation keeps the infrastructure simple and cost-effective, allowing you to focus on scaling volume rather than managing individual voice profiles.

  • No custom voice cloning offered
  • Consistent preset quality for ML training
  • Licensed source material reduces legal overhead

Comparison with SSML-based generators

SynthEval does not support SSML, phoneme dictionaries, or custom pronunciation lexicons. We prioritize raw text input and speed adjustment (0.25–4.0) for simplicity. This lack of granular control is a trade-off for easier integration with standard OpenAI SDKs and faster processing times for bulk generation tasks.

Most SSML-heavy APIs require complex markup parsing. Our approach reduces the engineering overhead for startups integrating audio generation into their data pipelines, allowing direct text-to-audio conversion without additional preprocessing steps.

  • No SSML support
  • No phoneme dictionaries
  • Simple text input only

Contrast with streaming-only APIs

We do not offer WebSocket or real-time streaming endpoints. Audio files are returned as complete responses after synthesis. This batch-oriented model is more efficient for generating large training datasets where latency per second is less critical than throughput and cost efficiency.

Streaming APIs often incur higher costs for short utterances. Our character-based pricing ($0.01 per 1,000 characters) and flat rate structure provide predictable costs for bulk audio generation, making it easier to budget for ML projects.

  • No WebSocket support
  • Batch audio delivery
  • Predictable character-based pricing

Technical Features for Developers

Sound wave illustration for: Technical Features for Developers

SynthEval integrates seamlessly into existing ML pipelines by mirroring the OpenAI speech endpoint structure. Developers can reuse official OpenAI SDKs simply by updating the base_url to https://api.play98f.com/v1, requiring minimal code changes for adoption.

OpenAI SDK compatibility

The API accepts standard JSON bodies identical to OpenAI's speech endpoint, supporting models tts-1 and tts-1-hd. This compatibility allows teams to switch providers without rewriting their inference logic, ensuring a smooth transition for existing infrastructure.

  • Supports tts-1 and tts-1-hd models
  • Identical request shape to OpenAI
  • Uses standard Bearer token authentication

Support for multiple audio formats

Generate audio in mp3, wav, opus, flac, or pcm formats directly via the response_format parameter. This flexibility ensures compatibility with various speech recognition models and storage requirements without needing external conversion tools.

  • Output formats: mp3, wav, opus, flac, pcm
  • Direct format selection in API request
  • No external conversion needed

Rate limits and queue management

Keys are limited to 60 requests per minute (10 per minute before the first top-up). Requests process sequentially, so high-volume dataset generation may queue during peak load, requiring patience for large batches.

  • 60 requests/minute per key
  • 10 requests/minute for new accounts
  • Sequential processing may cause queuing

Language and Voice Coverage

Sound wave illustration for: Language and Voice Coverage

SynthEval provides 23 languages including Arabic, Chinese, Japanese, and Swahili, all accessible through every voice preset. The system automatically detects script-based language for Latin text, but explicit ISO codes are recommended for non-English Latin inputs to ensure accuracy. Unsupported languages like Vietnamese and Ukrainian are clearly excluded.

The eight voice presets cover distinct personalities: female voices like Aria and Luna for narration, and male voices like Felix and Atlas for news. Each voice is derived from licensed LibriTTS-R and LibriVox recordings, ensuring clear, synthetic speech suitable for training data without licensing ambiguity.

23 languages supported across all voices

The API supports 23 languages, allowing any of the eight voices to speak any supported language. If the language parameter is omitted, the system guesses from the script, defaulting to English for Latin text. Explicitly setting the language code prevents misidentification for non-English Latin scripts.

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

8 distinct voice personalities

Eight presets offer varied tones for different dataset needs. Female voices include Aria (warm narrator), Nova (upbeat), Luna (storytelling), and Sage (instructional). Male options feature Felix (energetic), Milo (conversational), Atlas (neutral news), and Orion (deep documentary). All are based on licensed public-domain recordings.

Speed adjustment capabilities

Adjust the playback speed between 0.25 and 4.0 times normal rate. This flexibility helps generate diverse audio patterns for model training. Requests are processed sequentially, so queue times may vary under load, but typical generation speed approximates real-time audio duration.

Pricing and Cost Efficiency

Sound wave illustration for: Pricing and Cost Efficiency

SynthEval charges a flat rate of $10 per million characters, equating to $0.01 per 1,000 characters. Spaces and punctuation count toward this total. Failed requests incur no charge, ensuring you pay only for successful synthesis. This transparent model eliminates hidden fees for startups scaling dataset generation.

The absence of monthly subscriptions or seat licenses means you only pay for what you consume. There is no minimum spend or tiered pricing structure. This pay-as-you-go approach aligns costs directly with your data production volume, making budgeting predictable for variable workloads.

Flat rate per million characters

Each character in your input text contributes to the bill. At $0.01 per 1,000 characters, generating one hour of speech typically costs a fraction of a cent depending on speed. This linear pricing model simplifies cost estimation for large-scale ML training runs.

  • $10 per 1,000,000 characters
  • $0.01 per 1,000 characters
  • Failed requests are free

No monthly subscriptions or seats

Unlike platform-based competitors, SynthEval does not require recurring monthly fees or per-user licenses. You can start with the $0.50 trial credit or proceed directly to top-ups. This structure is ideal for project-based funding or irregular usage patterns common in early-stage AI ventures.

USDT payment flexibility

Top-ups are accepted via USDT on the TRON network (TRC-20), with a $5 minimum. Credits are automatic and balance does not expire. No cards, PayPal, or VAT invoices are used.

Data Privacy and Security

Sound wave illustration for: Data Privacy and Security

SynthEval processes input text solely to generate the requested audio file, ensuring your proprietary corpus remains private. Unlike many cloud TTS providers, we do not use your submitted text for model training or fine-tuning, preserving the exclusivity of your linguistic data for internal ML pipelines.

Our minimal data retention policy keeps only request metadata—such as timestamp, character count, and status—for billing verification. The generated audio is returned directly in the response and is not stored on our servers, reducing long-term storage liabilities and simplifying compliance audits for sensitive datasets.

Input text not used for training

Your input text is used exclusively for synthesis. We do not retain it for retraining our engine, ensuring that your unique phrasing or proprietary terminology remains distinct from our general training data. This separation guarantees that your dataset’s intellectual property is not inadvertently absorbed into the base model.

  • Text is processed in real-time for audio generation only.
  • No retention of input text for model improvement.
  • Preserves exclusivity of proprietary linguistic data.

Minimal data retention policies

We retain only essential metadata—time, character count, and status—for billing purposes. Generated audio files are not stored on our infrastructure after delivery, minimizing storage costs and reducing the attack surface for data breaches. This lean approach simplifies data governance for startups managing large-scale dataset volumes.

  • Metadata kept for billing verification only.
  • Audio files are not stored post-delivery.
  • Reduced storage footprint for large volumes.

No third-party data sharing

Your data stays within our direct processing pipeline. We do not share your input text or generated audio with third-party vendors, advertisers, or analytics providers. This ensures complete control over your dataset’s lineage, which is critical for maintaining audit trails in regulated AI development environments.

  • No sharing with third-party vendors.
  • No use for advertising or analytics.
  • Complete control over dataset lineage.
Data Privacy and Security

Use Case: Bias Reduction in AI

Sound wave illustration for: Use Case: Bias Reduction in AI

SynthEval enables developers to generate synthetic speech datasets with precise control over voice characteristics. By leveraging eight distinct voice presets, teams can create diverse speaker profiles without the licensing overhead of traditional voice actors.

Generating diverse speaker profiles

Use SynthEval’s eight voices—such as the warm, clear narrator Aria or the neutral news reader Atlas—to construct balanced audio datasets. Each voice supports 23 languages, allowing for consistent cross-lingual bias testing. The inaudible watermark ensures generated audio is clearly identified as synthetic, maintaining dataset integrity.

  • Leverage 8 distinct voice presets for varied tonal qualities
  • Support 23 languages to test cross-lingual model fairness
  • Use inaudible watermarks to label synthetic data accurately

Creating balanced demographic datasets

Generate labeled audio at scale using character-based pricing ($0.01 per 1,000 characters). This model allows startups to produce large volumes of speech data without monthly subscriptions. The consistent engine ensures uniform audio quality across all generated samples, reducing variance in training data.

  • Flat rate pricing allows predictable budgeting for large datasets
  • No monthly subscriptions reduce overhead for experimental projects
  • Uniform engine quality ensures consistent training data

Auditing voice recognition fairness

Test speech-to-text models against diverse voices to identify recognition biases. SynthEval’s API returns audio files directly, enabling automated evaluation pipelines. By varying speed and format, you can assess model robustness across different acoustic conditions without needing human recording studios.

  • Automate bias testing with direct audio file responses
  • Adjust speed (0.25–4.0x) to test model robustness
  • Export in MP3, WAV, or FLAC for standard pipelines

Use Case: Low-Resource Languages

Sound wave illustration for: Use Case: Low-Resource Languages

SynthEval addresses the scarcity of high-quality audio for less widely spoken languages by providing consistent, synthetic voices across 23 supported languages. This capability allows ML teams to generate labeled datasets for languages that lack extensive proprietary training data, ensuring model robustness across diverse linguistic inputs.

The API supports a wide range of voices speaking any of the 23 languages, enabling the creation of balanced datasets without the need for native speakers for every recording session. This flexibility is crucial for developing speech recognition models that perform well in multilingual environments.

Filling gaps in speech datasets

Many low-resource languages suffer from a lack of diverse, high-quality audio samples. SynthEval allows developers to generate thousands of unique utterances in these languages, filling critical gaps in training data. This is particularly useful for languages with limited native speaker populations or regional dialects that are underrepresented in major datasets.

By generating synthetic speech, teams can create varied prosody and intonation patterns, improving the model's ability to handle natural speech variations. This approach reduces the dependency on expensive, time-consuming human recording sessions for languages with sparse data availability.

  • Generate diverse utterances for underrepresented languages
  • Create balanced datasets with varied prosody
  • Reduce dependency on native speaker recordings

Supporting underrepresented languages

SynthEval supports 23 languages, including several that are often overlooked in mainstream TTS solutions. This includes languages like Swahili, Greek, and Turkish, which are increasingly important for global AI applications. The ability to specify language codes ensures accurate pronunciation and tone, even for languages with complex scripts.

The consistent voice quality across all supported languages ensures that models trained on synthetic data perform reliably. This is essential for applications like voice assistants and translation services that need to support a wide range of linguistic inputs without significant performance degradation.

  • Support for 23 diverse languages
  • Accurate pronunciation via language codes
  • Consistent voice quality across languages

Cost-effective scaling for niche markets

With a flat rate of $10 per million characters, SynthEval offers a cost-effective solution for scaling dataset generation. This pricing model eliminates the need for expensive per-voice licensing or subscription fees, making it feasible for startups to generate large volumes of audio data for niche markets.

The ability to generate audio on demand allows teams to scale their dataset size up or down based on specific project needs. This flexibility is particularly valuable for prototyping and iterative model development, where the required dataset size may vary significantly between experiments.

  • Flat rate pricing for scalable generation
  • No subscription fees for niche markets
  • On-demand audio generation for prototyping

Integration Workflow

Sound wave illustration for: Integration Workflow

To begin integration, navigate to /get-api-key/ on play98f.com and authenticate via Google or email. The dashboard displays your active API key, current balance, and a ready-to-use cURL example. This key authenticates all requests to the API base https://api.play98f.com/v1. Ensure you copy the key starting with tts-... for immediate use.

Step-by-step API key retrieval

Register an account on the /get-api-key/ page using Google or email credentials. The interface immediately provides your unique API key, current balance, and usage statistics. This key authorizes your requests to https://api.play98f.com/v1.

  • Sign up via Google or email on /get-api-key/
  • Retrieve your API key starting with 'tts-...'
  • Copy the key for your application configuration

Sending first speech request

Construct a POST request to /v1/audio/speech with a JSON body. Include the model (tts-1 or tts-1-hd), input text (max 4096 characters), and voice preset. Set the Authorization header with your Bearer key. The API returns the audio file directly in the response body.

Use the OpenAI SDK by setting base_url to https://api.play98f.com/v1. This allows you to use standard client libraries without custom implementation. Ensure your text respects the 4096-character limit per request to avoid truncation.

  • Set base_url to https://api.play98f.com/v1
  • Include model, input, voice, and authorization headers
  • Ensure input text is under 4096 characters

Handling responses in Python

Parse the HTTP response to extract the audio bytes. Save the file using the specified response_format (mp3, wav, opus, flac, or pcm). The SDK handles the binary data automatically, allowing direct writing to disk or memory buffers for further processing in your ML pipeline.

No additional parsing is required for the audio content itself. Simply write the response content to a file with the correct extension. This streamlined process ensures quick integration into your dataset generation scripts.

Common Mistakes to Avoid

Sound wave illustration for: Common Mistakes to Avoid

Many developers lose time and credits by sending raw strings exceeding 4096 characters, which the API rejects outright. To handle bulk generation, split text into chunks before submission and concatenate the resulting audio files. This ensures consistent delivery without request failures.

Ignoring character limits in bulk generation

The API enforces a strict 4096-character limit per request. Sending larger texts results in immediate rejection, wasting a request unit. Developers must implement client-side splitting logic to join multiple audio segments seamlessly.

Not setting language codes for Latin script

If the language field is omitted, the system defaults to English for Latin-script text. This causes incorrect pronunciation for non-English languages like French or German using the same Latin alphabet. Always specify the ISO code to ensure accurate synthesis.

Assuming SSML support where it does not exist

Unlike some competitors, SynthEval does not support SSML tags for prosody or pronunciation. Custom lexicons and emotion tags are also absent. Rely on the eight preset voices and speed adjustments rather than expecting granular XML-based control.

Common Mistakes to Avoid

Final Recommendations

Sound wave illustration for: Final Recommendations

SynthEval provides a transparent, watermark-marked TTS API ideal for startups needing scalable audio datasets without licensing overhead. Its OpenAI-compatible SDK and flat character-based pricing simplify integration, while the lack of streaming or SSML keeps the tool focused on batch generation for training pipelines.

Best choice for ML dataset creators

SynthEval is a strong tool for generating labeled audio datasets at scale. The use of LibriTTS-R voices and clear licensing removes ambiguity, while the inaudible watermark ensures synthetic origin is always identifiable for model training.

Ideal for startups needing scalable audio

Startups benefit from the simple USDT top-up model and no monthly subscriptions. The API handles bulk requests efficiently, allowing teams to generate thousands of audio files for diverse models without managing complex seat licenses or enterprise contracts.

Recommended for bias auditing projects

Generating diverse speaker profiles across 23 languages helps audit voice recognition fairness. The consistent voice presets allow for controlled experiments, ensuring that demographic biases in speech recognition systems are identified and corrected systematically.

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. OpenAI API reference: create speech
  3. LibriTTS-R corpus (OpenSLR 141)
  4. IETF RFC 7845: Ogg Opus
  5. Python documentation
  6. Creative Commons Attribution 4.0
  7. MDN: Web Speech API
  8. openai-python on GitHub
  9. ISO 639 language codes
  10. curl documentation
  11. LibriVox public-domain audiobooks
  12. Tether (USDT)

More on this site