Frequently asked questions
What is synthetic speech in ML?
Synthetic speech is audio generated by a TTS API, such as SynthEval, using text input. It provides labeled audio datasets for training machine learning models. The API returns audio files directly, enabling startups to scale data generation without licensing overhead or manual recording sessions.
How is synthetic speech different from TTS?
Synthetic speech is the output of a TTS engine. While TTS refers to the technology converting text to audio, synthetic speech specifically denotes the generated audio files used as training data. Our API delivers these files via a standard endpoint, optimized for dataset creation rather than end-user playback.
Can synthetic voices replace human recorders?
Synthetic voices offer a scalable alternative for generating large datasets. They eliminate the need for studio time and human talent. However, they lack human nuance. For specific emotional or contextual accuracy, human recorders may still be required. Synthetic speech is effective for volume and consistency in training data.
What are the licensing terms for synthetic audio?
Our voices are based on licensed recordings from LibriTTS-R (CC BY 4.0) and LibriVox public-domain readings. Generated audio carries an inaudible watermark marking it as synthetic. Input text is not used for training. This structure helps reduce licensing overhead when integrating synthetic audio into your machine learning pipelines.
How do watermarks help in ML datasets?
Every generated file carries an inaudible watermark marking it as synthetic speech. This allows you to distinguish synthetic data from human-recorded audio in your datasets. It ensures transparency and helps in training models to recognize synthetic sources, which is critical for evaluating model performance on mixed data types.
Is synthetic speech accurate enough for training?
Synthetic speech provides high-quality audio for training pipelines. Voices are based on licensed recordings, ensuring clear and natural sound. While it may not match every human nuance, it is effective for scaling labeled datasets. The API supports multiple languages and voices, allowing you to test accuracy for your specific use case.
What languages are supported for synthetic data?
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. Other languages are not supported.
How do I integrate TTS into my ML pipeline?
Send text to our endpoint via POST request. Use the standard OpenAI SDK by changing the base URL to https://api.play98f.com/v1. Provide your API key in the Authorization header. Receive audio files directly in your pipeline. This standard integration allows seamless automation of dataset generation for your machine learning workflows.
Does synthetic data introduce bias?
Synthetic data reflects the voices and languages provided by the API. Bias depends on the diversity of voices and languages used in your dataset. We offer eight distinct voices across genders. You can control bias by selecting specific voices and languages, ensuring your training data represents your target demographic accurately.
What formats are best for ML audio inputs?
We support mp3, wav, opus, flac, and pcm formats. For ML training, uncompressed formats like wav or pcm are often preferred for quality. Encoded formats like mp3 or opus reduce storage costs. Choose the format that best matches your model’s input requirements and your storage efficiency needs.
How much data do I need for training?
Data requirements depend on your specific model and task. Our API allows generating up to 4096 characters per request. You can scale by making multiple requests. With a rate limit of 60 requests per minute, you can generate significant volumes of data efficiently to train robust machine learning models.
Can I use synthetic data for commercial products?
Yes, you can use synthetic audio for commercial products. The voices are based on licensed recordings. Generated audio is returned in the response and not stored by us. You pay for usage, and failed requests are free. This model supports commercial scaling without ongoing licensing fees for the synthetic content.
What is the difference between tts-1 and tts-1-hd?
Both models use the same engine and price. tts-1-hd offers higher fidelity. tts-1 is suitable for general use. Choose based on your quality needs and budget, as the cost per character is identical.
Do I need to label synthetic audio manually?
No, you do not need to label synthetic audio manually. The API generates audio from text input, so you already have the corresponding text label. This automation reduces the labor required for dataset creation. You can directly pair the generated audio with its source text for training your machine learning models.
Definition of Synthetic Speech
Synthetic speech is audio generated by algorithms from text input, rather than recorded by human actors. This technology uses neural networks to synthesize natural-sounding voices, enabling the creation of vast audio datasets without the logistical overhead of studio sessions.
Text-to-Speech (TTS) serves as a critical component in modern data pipelines. By converting written text into audio, TTS allows ML engineers to generate labeled training data at scale, ensuring diverse acoustic variations for robust model training.
Audio generated by algorithms vs human recording
Unlike human recordings which capture unique environmental nuances and emotional imperfections, algorithmic speech is produced deterministically from text. Each output is a fresh synthesis, ensuring consistency across millions of generated samples while remaining distinct from the source voice data.
- Deterministic generation from text input
- Consistent acoustic properties across samples
- No physical recording constraints
Role of Text to Speech in data pipelines
TTS integrates directly into ML workflows via API endpoints. Engineers send text payloads and receive audio files, automating the creation of labeled datasets. This integration supports high-volume generation, essential for training speech recognition models that require diverse phonetic coverage.
- API-driven data generation
- Automated labeling workflows
- Scalable dataset creation
Difference between TTS and voice cloning
TTS synthesizes speech from a predefined set of voices, whereas voice cloning requires training a model on specific voice samples. Our service uses preset voices derived from licensed recordings, offering variety without the need for custom voice training or cloning infrastructure.
Voice cloning typically demands extensive data and computational resources to replicate a specific individual's timbre and style. In contrast, standard TTS provides immediate access to distinct voices like Aria or Felix, suitable for general-purpose synthetic data generation.
- Predefined voice presets
- No custom model training required
- Immediate availability
Applications in ML Training
Synthetic speech enables efficient augmentation of limited speech datasets by generating unlimited text-to-audio pairs. By leveraging a robust TTS API, ML teams can rapidly expand training corpora without the logistical overhead of human recording studios. This approach ensures that models are exposed to a wide variety of acoustic conditions and phonetic combinations.
Augmenting small speech datasets
Generating labeled audio datasets at scale allows startups to train models on niche domains with minimal initial data. The API processes text up to 4096 characters, enabling the creation of long-form training samples or segmented phrases. This scalability reduces the cost per hour of audio significantly compared to manual narration.
- Scale datasets from thousands to millions of hours efficiently
- Reduce licensing overhead for commercial training data
- Maintain consistent audio quality across all generated samples
Generating diverse accents and tones
With eight distinct voice presets spanning male and female ranges, synthetic speech provides immediate diversity. Each voice can speak any of the 23 supported languages, allowing for precise control over tonal variations. This variety helps model generalization across different speaker profiles without recruiting numerous voice actors.
- Eight presets: from warm narrators to energetic promo voices
- Support for 23 languages with optional ISO code specification
- Consistent character-level pricing for diverse voice generation
Stress testing voice recognition systems
Synthetic audio is ideal for stress-testing speech-to-text engines under controlled conditions. You can generate edge-case inputs, such as specific speeds or language mixes, to identify model weaknesses. The inaudible watermark also helps verify that the system correctly identifies synthetic vs. human speech when needed.
- Test edge cases with precise speed and language parameters
- Verify model robustness against algorithmic acoustic signatures
- Automate generation pipelines for continuous regression testing

Data Quality and Licensing
Synthetic speech datasets require clear provenance to ensure model robustness. Our voices are derived from the LibriTTS-R dataset, built upon public-domain recordings from LibriVox. This foundation provides a transparent lineage for your training data, avoiding the ambiguity often associated with scraped web audio.
Understanding the licensing structure is critical for commercial ML products. The underlying recordings are released under CC BY 4.0, which permits commercial use with attribution. This clarity removes legal friction when integrating generated audio into proprietary speech recognition pipelines.
Each generated audio file contains an inaudible watermark marking it as synthetic. This allows your evaluation pipelines to distinguish generated data from human recordings, ensuring accurate bias audits and preventing synthetic data leakage during model training phases.
Source of voices: LibriTTS-R and LibriVox
All eight voice presets are based on licensed recordings from the LibriTTS-R dataset, which aggregates public-domain audiobook readings from LibriVox. This ensures that the acoustic characteristics are rooted in real human speech while maintaining a clear, auditable origin for your datasets.
Using established, publicly documented sources simplifies compliance checks. Since the base recordings are already released under open licenses, you avoid the complex rights management often required when scraping or licensing individual voice actors for large-scale data generation.
- Rooted in LibriVox public-domain audiobooks
- Curated into the LibriTTS-R dataset
- Transparent acoustic lineage for ML training
CC BY 4.0 licensing implications
The CC BY 4.0 license allows commercial use of the underlying voice data, provided appropriate credit is given. This means you can freely integrate the generated synthetic audio into commercial ML models without worrying about restrictive royalty structures or per-unit fees for the voice IP itself.
This licensing model is particularly advantageous for startups generating millions of audio samples. You retain full ownership of the synthetic dataset you produce, while the base voice licenses remain permissive, reducing legal overhead during scaling phases.
- Commercial use permitted
- Attribution required
- No per-unit voice royalties
Inaudible watermarks for synthetic detection
Every audio file generated via our API carries an inaudible watermark that identifies it as synthetic speech. This feature is essential for maintaining data integrity in mixed training sets, allowing your models to learn the distinct acoustic signature of generated audio.
You can use these watermarks during evaluation to measure how well your models generalize from synthetic to real speech. This helps quantify the 'sim-to-real' gap and ensures your final deployed models perform reliably on human voice data.
- Inaudible synthetic markers
- Enables dataset filtering
- Supports bias auditing
Advantages Over Human Recording
Synthetic speech eliminates the logistical bottlenecks of human recording studios. By leveraging an API, you can generate millions of hours of audio on demand without scheduling session time or managing studio overhead. This allows startups to scale dataset creation linearly with computational resources rather than human availability.
Scalability to millions of hours
Unlike human voice actors limited by working hours and stamina, the API processes requests sequentially with predictable throughput. You can generate terabytes of labeled audio by chaining requests, turning data collection from a manual project into an automated pipeline that runs 24/7.
- No studio scheduling conflicts
- Linear scaling with API keys
- Continuous generation without fatigue
Consistent audio quality
Every generated file maintains uniform acoustic properties, ensuring your ML models train on consistent signal-to-noise ratios. There are no variations in microphone distance, room tone, or speaker health between takes, which reduces the need for extensive post-processing and normalization in your training pipeline.
- Uniform background noise profile
- Stable volume and frequency response
- No vocal fatigue variations
Lower cost per hour of audio
At $0.01 per 1,000 characters, synthetic audio is significantly cheaper than hiring voice talent for large-scale datasets. For large-scale ML datasets, this pricing model reduces the marginal cost of additional training data to near zero, allowing you to experiment with more diverse text samples without budget constraints.
- $0.01 per 1k characters
- No hourly studio rates
- Predictable scaling costs
Common Use Cases
Synthetic speech enables efficient data generation for machine learning pipelines. By producing labeled audio at scale, developers can train models without the licensing overhead of human recordings.
Training speech-to-text models
Generate diverse accents and tones to augment small speech datasets. This approach helps stress-test voice recognition systems under varied conditions.
- Augment limited training data with synthetic variations
- Stress-test recognition accuracy across different voices
Creating synthetic speakers for bias audits
Create balanced speaker sets to identify algorithmic bias. Synthetic voices allow for controlled experiments on demographic representation without recruiting human subjects.
- Ensure balanced representation in test sets
- Identify bias without human subject recruitment
Generating audio for low-resource languages
Produce audio for languages with scarce data. Our API supports 23 languages, enabling training for models that lack sufficient real-world recordings.
- Support for 23 languages
- Enable training for low-resource language models
Technical Requirements
Integrating a TTS API into ML pipelines requires handling HTTP requests and parsing JSON responses. Using the OpenAI SDK with a custom base_url simplifies this, allowing developers to fetch audio files directly into training datasets without manual file management.
The API accepts inputs up to 4,096 characters per request. Scripts exceeding this limit must be split into parts, and the resulting audio files concatenated to preserve continuity. This character-based billing model means precise script management is essential for cost control.
Need for API integration in pipelines
Automating data generation requires robust API calls. Developers typically use the OpenAI SDK, changing the base_url to https://api.play98f.com/v1. This allows seamless integration into Python-based workflows, where text inputs are sent and audio bytes are received for immediate storage.
The API supports standard HTTP methods, making it compatible with most pipeline tools. Each request returns a distinct audio file, which can be programmatically labeled with metadata like voice type and language for structured dataset creation.
- Use OpenAI SDK with custom base_url
- Automate JSON request/response handling
- Supports Python and standard HTTP clients
Handling audio formats (wav, mp3, opus)
The API outputs audio in multiple formats: mp3, wav, opus, flac, and pcm. For ML training, uncompressed formats like wav or pcm are often preferred for quality, while opus offers efficient compression. The response_format parameter in the JSON body specifies the desired output.
Each generated file includes an inaudible watermark to identify synthetic speech. This helps in curating datasets by distinguishing AI-generated audio from human recordings, ensuring transparency in training data composition.
- Supported formats: mp3, wav, opus, flac, pcm
- Uncompressed wav/pcm for high-quality training
- Inaudible watermark for synthetic detection
Managing character limits in scripts
Each API request is limited to 4,096 characters. For longer scripts, developers must split the text into smaller segments, generate audio for each, and concatenate the results. This ensures consistent audio quality and avoids truncation.
Spaces and punctuation count toward the character limit. Accurate script parsing is crucial to avoid unexpected cuts or billing issues. The API processes requests sequentially, so large batches may queue, requiring careful timing in automated pipelines.
- Max 4,096 characters per request
- Split and concatenate for long texts
- Spaces and punctuation count as characters
Limitations of Synthetic Data
Synthetic speech models, while highly effective for scaling datasets, inherently lack the subtle emotional nuances present in human recordings. Our voices are based on licensed recordings from LibriTTS-R, providing clear narration but missing the complex prosody of spontaneous human interaction.
Lack of human emotion nuances
Our voices are based on licensed recordings from LibriTTS-R, providing clear narration but missing the complex prosody of spontaneous human interaction. This can limit suitability for training models where emotional inflection is critical.
Potential for algorithmic bias
Since our voices derive from specific source datasets, they may reflect the demographic or linguistic characteristics of those recordings. This can introduce subtle biases if the training data does not adequately represent diverse speaker profiles.
Distinct acoustic signature from real speech
Every generated file carries an inaudible watermark that marks it as synthetic speech. This distinct acoustic signature helps identify synthetic data but may require specific handling during model training to ensure accurate classification.
Integration with ML Frameworks
Synthetic speech APIs integrate directly into ML pipelines by returning raw audio bytes, which can be immediately decoded into tensors for TensorFlow or PyTorch. This eliminates the need for intermediate file storage, allowing seamless ingestion of generated audio into training datasets for speech-to-text or speaker verification models.
The API’s support for multiple formats like WAV, MP3, and OPUS ensures compatibility with standard data loaders. By fetching audio via HTTP requests, teams can programmatically generate diverse acoustic conditions and speaker variations on demand, maintaining consistent data quality without manual recording overhead.
Using API responses in TensorFlow/PyTorch
Receive audio bytes via HTTP and decode them into numpy arrays. Pass these directly to model input layers, skipping disk I/O for faster iteration cycles during hyperparameter tuning.
Support for WAV and PCM simplifies integration with Keras and PyTorch audio libraries. The synchronous nature of the API allows predictable batching for training runs.
- Direct byte-to-tensor conversion
- No intermediate file storage required
- WAV/PCM compatibility with standard loaders
Automating data generation workflows
Use the API’s character limit (4096 chars) to script batched generation. Combine multiple short texts or split longer scripts to create diverse audio samples automatically.
Integrate with cron jobs or CI/CD pipelines to regenerate datasets with updated voice presets or languages, ensuring your model trains on the latest synthetic data variations.
- Scripted batch generation
- CI/CD pipeline integration
- Dynamic voice switching
Monitoring generation quality metrics
Track request status and character counts via the API’s response metadata. Use this data to audit dataset composition, ensuring balanced representation across voices and languages.
Monitor generation speed and failure rates to optimize pipeline throughput. Since failed requests are free, you can safely retry low-confidence generations without cost penalties.
- Request status tracking
- Character count auditing
- Retry logic for failures
Future of Synthetic Audio
Synthetic speech generation is evolving from rule-based systems to advanced neural architectures that produce more natural prosody and reduced acoustic artifacts. As models mature, the distinction between human and machine-generated audio becomes increasingly subtle, enabling higher fidelity datasets for training. This progression supports more robust machine learning pipelines by providing diverse, high-quality audio samples without the logistical constraints of human recording sessions.
The technology is gaining traction in enterprise AI as companies seek scalable solutions for voice recognition and synthetic speaker data. Organizations are leveraging these APIs to rapidly prototype and deploy voice-enabled products.
Advancements in neural TTS models
Neural TTS models are improving in naturalness and expressiveness, allowing synthetic voices to mimic human-like intonation more accurately. These advancements enable the generation of longer, coherent audio segments, reducing the need for complex stitching algorithms in data pipelines.
Current models like 'tts-1-hd' provide high-fidelity output. This reduces the risk of model bias when training speech recognition systems on mixed datasets.
- Improved prosody and intonation
- Reduced acoustic artifacts
- Higher fidelity output options
Growing adoption in enterprise AI
Enterprises are increasingly adopting synthetic speech for training voice assistants, call center bots, and accessibility tools. The ability to generate diverse accents and tones at scale allows companies to test their systems against a wider range of user demographics, improving overall performance.
Synthetic data generation is becoming a standard practice in AI development, enabling rapid iteration and testing. Companies are using TTS APIs to create labeled datasets for speech-to-text models, ensuring consistent quality and reducing dependency on human voice actors.
- Voice assistant training
- Call center bot development
- Accessibility tool testing
Standardization of synthetic data labels
As synthetic speech becomes more prevalent, there is a growing need for standardized labeling to distinguish synthetic audio from human recordings. Inaudible watermarks, such as those used in our generated files, provide a reliable method for identifying synthetic content in large datasets.
Standardized labels help ensure transparency and compliance in AI training data. By clearly marking synthetic audio, organizations can maintain audit trails and verify the source of their training data.
- Inaudible watermarks
- Data source transparency
- Regulatory compliance
Choosing a TTS Provider
Selecting a TTS provider for ML pipelines requires balancing voice diversity with technical reliability. The SynthEval API offers eight distinct presets—ranging from warm narrators to deep documentary tones—capable of rendering 23 languages. This variety allows teams to generate diverse acoustic profiles for bias audits and low-resource language training without licensing overhead.
Reliability hinges on predictable performance and clear cost structures. Providers should offer consistent audio formats like WAV or FLAC for raw data ingestion. Evaluate whether the API handles character limits gracefully and maintains stable throughput under load, ensuring your training scripts receive consistent, high-quality synthetic audio without unexpected downtime or variable quality.
Evaluating voice quality and variety
Ensure the provider offers a sufficient range of voices and languages. SynthEval provides eight presets across 23 languages, all based on licensed recordings. This variety supports the creation of diverse datasets, crucial for training robust speech recognition models that generalize well across different acoustic environments and speaker characteristics.
- Check for multiple distinct voice personalities (e.g., warm, neutral, deep).
- Verify support for the specific languages required in your dataset.
- Ensure voices are distinct enough to prevent model overfitting to a single timbre.
Checking licensing and usage rights
For ML training data, licensing clarity is critical. SynthEval uses voices from LibriTTS-R (CC BY 4.0) and LibriVox, simplifying rights management. Additionally, every generated file includes an inaudible watermark, allowing you to easily distinguish synthetic data from human recordings in your training sets, which is vital for unbiased model evaluation.
- Confirm the license allows commercial use for training AI models.
- Check if watermarks are embedded for synthetic data detection.
- Verify that input text is not used for training other models.
Assessing API reliability and speed
High-volume generation demands a stable API. SynthEval handles requests sequentially, with speeds roughly matching real-time generation for long texts. With a limit of 60 requests per minute, plan your batch jobs accordingly. Ensure the API returns standard formats like WAV or MP3, which integrate seamlessly into TensorFlow or PyTorch data loaders without extra conversion steps.
- Verify request rate limits (e.g., 60 req/min) fit your pipeline throughput.
- Check supported output formats (WAV, MP3, FLAC) for direct integration.
- Test latency for short vs. long text inputs to estimate generation times.

Cost Analysis for ML Projects
Synthetic speech significantly reduces the cost per hour of audio compared to human recording, especially for large-scale ML datasets. With pricing at $0.01 per 1,000 characters, generating hours of training data becomes highly economical.
Budgeting for large-scale generation requires estimating total character volume across your dataset. Since failed requests are free, you can safely experiment with different voices and speeds without financial risk, optimizing your pipeline before full deployment.
Comparing cost per hour of audio
Human voice talent can cost $50-$100 per hour of finished audio. Synthetic speech costs roughly $0.01 per 1,000 characters, translating to pennies per hour of audio. This drastic reduction allows startups to train models on millions of hours without prohibitive labor costs.
The API processes requests sequentially, taking roughly 25 characters per second. For long texts, generation is near real-time. Short phrases may take a few seconds, but the low per-request cost makes this efficient for batch processing.
- Human recording: ~$50-$100/hour
- Synthetic speech: ~$0.001-$0.01/hour (depending on text density)
- Failed requests are free
- No minimum spend or monthly fees
Budgeting for large-scale generation
Start with the free trial credit of $0.50 (50,000 characters) to validate your pipeline. For production, top up with USDT (TRC-20) starting at $5. Balance does not expire, allowing flexible budgeting without subscription lock-ins.
Rate limits allow 60 requests per minute. Plan your generation jobs to stay within this limit, or queue longer texts. Split texts into 4096-character chunks to ensure consistent quality and manageable processing times.
- Free trial: $0.50 credit
- Top-up: USDT (TRC-20), min $5
- Rate limit: 60 req/min
- No expiration on balance
Impact of character limits on pricing
Pricing is based on character count, including spaces and punctuation. Each request supports up to 4096 characters. Splitting long scripts into multiple requests allows fine-grained control over audio segments and easier debugging.
Longer texts are split and joined automatically. Ensure your pipeline handles the resulting audio files correctly. The consistent pricing model means you can accurately predict costs based on your dataset size before generation begins.
- Max 4096 characters per request
- Spaces and punctuation count
- Split texts for control
- Predictable cost per character
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.
23 languages
Any voice speaks any of these languages.
Sources and further reading
- Wikipedia
- curl documentation
- openai-node on GitHub
- FFmpeg documentation
- W3C WAI: making audio and video accessible
- MDN: Web Speech API
- Node.js documentation
- LibriTTS-R corpus (OpenSLR 141)
- LibriVox public-domain audiobooks
- Python documentation
- Creative Commons Attribution 4.0
- LibriTTS-R paper (arXiv 2305.18802)
