Frequently asked questions
What makes a TTS API good for ML datasets?
A good TTS API for ML datasets offers low cost-per-character, broad language coverage, and clear data rights. Our API charges $0.01 per 1,000 characters and supports 23 languages. Input text is not used for training, and generated audio is not stored, ensuring clean, isolated data for your models without leakage or licensing ambiguities.
Are the voices licensed for commercial use?
Yes. The voices are based on licensed recordings from LibriTTS-R (CC BY 4.0) and public-domain audiobook readings from LibriVox. This provides a solid foundation for commercial use in your synthetic speech datasets.
How diverse are the available voices?
We provide eight distinct voice presets: four female (aria, nova, luna, sage) and four male (felix, milo, atlas, orion). Each voice has a specific tonal quality, from warm narration to deep documentary styles, allowing you to create diverse audio samples for robust model training.
Can I generate audio in multiple languages?
Yes, we support 23 languages including English, Chinese, Spanish, and Hindi. Any voice can speak any supported language. You can optionally specify the ISO language code to ensure accuracy, especially for non-English Latin-script text, preventing incorrect language guessing by the system.
Is the API suitable for large-scale data generation?
Yes, with specific constraints. You can send up to 4,096 characters per request. For larger texts, split them into parts and join the audio files. Rate limits are 60 requests per minute per key, which requires batching strategies for massive dataset generation but remains cost-effective.
What audio formats are best for ML training?
We support mp3, wav, opus, flac, and pcm. For ML training, uncompressed formats like wav or pcm are often preferred to preserve audio fidelity without decoding artifacts. FLAC offers lossless compression to save storage space while maintaining exact audio quality for your datasets.
How much data can I generate with the free trial?
The free trial provides $0.50 of credit, valid for 7 days. At our rate of $0.01 per 1,000 characters, this allows you to generate approximately 50,000 characters of audio. This is sufficient for testing integration and evaluating voice quality before committing to paid top-ups.
Are there restrictions on using generated audio?
Generated audio carries an inaudible watermark marking it as synthetic speech. Input text is used only for production and not for training our models. There are no specific usage restrictions beyond the standard licensing of the source voice recordings, making it suitable for commercial ML datasets.
How does the API handle language detection?
If you do not specify a language, the system guesses it from the script. Latin script text defaults to English. For non-English Latin-script text, you must set the language parameter explicitly to ensure accurate synthesis, as the default guess may not match your intended language.
Can I use the API for speech recognition training?
Yes, it is ideal for speech recognition training. You can generate paired text-audio data by sending specific text inputs. The consistent character-to-audio ratio and support for multiple languages allow you to build diverse training sets for ASR models without needing human voice actors for every sample.
Is there a limit to the number of voices I can use?
No, there is no limit to the number of voices you can use. You can rotate through all eight presets (aria, nova, luna, sage, felix, milo, atlas, orion) across your requests. This flexibility helps you create varied acoustic profiles for your machine learning datasets.
How do I ensure data privacy with the API?
Input text is used solely to produce the audio and is not used for training our models. Generated audio is returned in the response and not stored on our servers. Only request metadata (time, characters, status) is kept for billing purposes, ensuring your raw data remains private.
Can I integrate the API with my existing ML pipeline?
Yes. The API follows the OpenAI speech endpoint shape, so official OpenAI SDKs work by changing the base URL. You can use standard HTTP requests or SDKs to send text and receive audio files, integrating seamlessly into existing data generation pipelines for synthetic speech production.
What is the turnaround time for large requests?
Typical speed is roughly real-time (25 characters per second). Short phrases take a few seconds. Requests are processed sequentially, so under load, a request may wait in a queue. For large datasets, plan for potential queuing delays during high traffic periods.
Defining Requirements for ML Datasets
Building robust speech datasets requires diverse acoustic profiles and broad linguistic coverage. This API offers eight distinct voice presets, ranging from warm narrators to energetic promo styles, providing the variety needed for model generalization without the need for manual data augmentation.
Voice Diversity Needs
The eight available voices—four female (aria, nova, luna, sage) and four male (felix, milo, atlas, orion)—cover different tonal ranges suitable for various ML training scenarios. This variety allows teams to create balanced datasets without relying on external voice actors.
- 8 distinct voice presets covering narration, conversation, and news styles
- Consistent audio quality across all voice types
Language Coverage Importance
Support for 23 languages enables the creation of multilingual models. The system automatically detects language from script but allows explicit specification, ensuring accuracy for non-English Latin text. Note that languages like Vietnamese, Thai, and Ukrainian are not supported.
- 23 languages including English, Chinese, Arabic, Hindi, and Spanish
- Automatic language detection with optional ISO code override
Cost Per Character Metrics
Pricing is straightforward at $10 per million characters, or $0.01 per 1,000 characters. This predictable cost structure aids in budgeting large-scale dataset generation, with failed requests refunded automatically.
- $0.01 per 1,000 characters
- Failed requests are free
Evaluating Voice Quality and Authenticity
Voice authenticity in synthetic datasets depends on source fidelity. Lucky TTS uses licensed LibriTTS-R recordings (CC BY 4.0) and public-domain LibriVox readings, providing a transparent, legally clear foundation for training pipelines. This ensures that generated audio retains the natural prosody of human speech without requiring additional rights clearance for the base recordings.
The synthetic nature is marked by an inaudible watermark, allowing downstream models to distinguish generated speech from real human recordings. This metadata is critical for curating balanced datasets where synthetic and real audio proportions are known and trackable.
Licensed Source Recordings
The eight voice presets (Aria, Nova, Luna, Sage, Felix, Milo, Atlas, Orion) are derived from LibriTTS-R (CC BY 4.0) and LibriVox public-domain audiobooks. This provides a clear, permissive license chain for commercial ML training, avoiding the ambiguity of voice-clone rights.
Users retain full control over the generated audio files. Since the source recordings are already licensed for reuse, the resulting synthetic audio is generally safe for commercial dataset construction, provided the specific model output is not considered a derivative work requiring additional attribution beyond the API terms.
- Source: LibriTTS-R (CC BY 4.0) and LibriVox public domain.
- Eight distinct voice presets based on licensed human recordings.
- Clear lineage for commercial ML dataset usage.
Synthetic Watermark Impact
Every generated audio file carries an inaudible digital watermark identifying it as synthetic speech. This allows ML pipelines to easily filter out or label synthetic samples, preventing data leakage between training and evaluation sets.
For dataset curation, this watermark ensures that synthetic data does not accidentally contaminate human-verified test sets, maintaining the integrity of model evaluation metrics.
- Inaudible watermark marks all outputs as synthetic.
- Facilitates precise data filtering and labeling.
- Prevents synthetic data leakage into test sets.
Noise and Artifacts
As a cloud-based TTS API, audio quality is consistent and free from studio noise, but may exhibit minor artifacts typical of neural speech generation. These are generally low-frequency and do not significantly impact most speech recognition or synthesis training tasks.
For high-fidelity requirements, users can select the 'tts-1-hd' model for enhanced clarity. The 'wav' or 'flac' response formats minimize compression artifacts, preserving the full spectral data for training models that require precise acoustic features.
- Consistent cloud-based audio quality.
- Select 'tts-1-hd' for enhanced clarity.
- Use 'wav' or 'flac' to minimize compression artifacts.
image":
Technical Compatibility with ML Tools
This API uses a standard OpenAI-compatible endpoint structure, sending POST requests to /v1/audio/speech with JSON bodies. This ensures seamless integration with existing ML data pipelines without requiring custom wrappers.
The request schema mirrors industry standards, accepting text input, voice selection, and output format parameters. This consistency simplifies batch processing scripts for generating large synthetic speech datasets.
Standard Endpoint Structure
The API exposes a single POST https://api.playlucky101.com/v1/audio/speech endpoint. Requests use JSON bodies identical to the OpenAI speech endpoint, allowing direct reuse of common HTTP clients and batch job scripts used in ML training pipelines.
This standardization reduces integration friction. Teams can pipe text files directly into the endpoint without adapting to proprietary request shapes, ensuring consistent data flow for model training.
- JSON body structure matches OpenAI standards.
- Single endpoint for all speech generation tasks.
- Supports standard HTTP methods for batch jobs.
SDK Flexibility
While there is no proprietary SDK, the endpoint compatibility allows the official OpenAI Python/Node SDKs to work by simply changing the base_url to https://api.playlucky101.com/v1. This provides immediate programmatic access for dataset generation scripts.
Developers can also use plain HTTP requests or any standard library. This flexibility is crucial for ML engineers who prefer lightweight, custom-built data loaders over heavy SDK dependencies.
This approach ensures that existing infrastructure for data ingestion remains intact, avoiding the need to rewrite core pipeline components just to accommodate a new TTS provider.
- Works with official OpenAI SDKs via base_url change.
- Plain HTTP requests fully supported.
- No custom SDK required for integration.
Format Support for Training
The API supports output formats including mp3, wav, opus, flac, and pcm. This variety allows teams to select formats based on their model's requirements, balancing file size and quality.
Raw PCM or WAV files are ideal for training speech recognition models, while compressed formats like MP3 or Opus reduce storage costs for large-scale datasets without significant quality loss.
Each generated file carries an inaudible watermark, marking it as synthetic. This aids in dataset curation by allowing easy identification of synthetic versus human-recorded data during model evaluation.
- Supports mp3, wav, opus, flac, and pcm.
- Raw formats ideal for ASR training.
- Inaudible watermark marks synthetic speech.
Pricing Models and Budgeting
For ML dataset pipelines, predictability matters. At $0.01 per 1,000 characters, costs scale linearly with corpus size. Failed requests incur no charge, protecting your budget during bulk generation phases where edge cases might trigger errors.
Pay-Per-Character Efficiency
The flat rate of $10 per million characters simplifies budgeting. Spaces and punctuation count, so accurate character counting prevents unexpected overages. This model works well for high-volume text normalization tasks.
Trial Credit Utilization
Google sign-ups grant $0.50 (approx. 50,000 characters) for 7 days. Use this to validate voice quality before committing funds. Email-only accounts start without this credit, so plan your initial testing accordingly.
USDT Payment Simplicity
Top-ups require USDT on TRON (TRC-20). Minimum $5 ensures low overhead for small batches. Credits appear automatically within minutes, avoiding invoice delays. Balance never expires, allowing you to store credits for future large-scale jobs.
Data Usage and Licensing Rights
The synthetic speech generated by this API is free for any use, including commercial ML datasets. The voice models originate from LibriTTS-R recordings licensed under CC BY 4.0, but the TTS engine synthesizes new audio from these base recordings.
This distinction simplifies data pipeline management. Since the generated audio is not a direct copy of a public-domain reading, you can mix this synthetic data with other sources without tracking individual attribution requirements for each voice. This reduces legal overhead when curating large-scale training sets for speech recognition or text-to-speech models.
CC BY 4.0 Licensed Source
The underlying voice models are based on LibriTTS-R, a dataset of public-domain audiobook readings. While the source recordings are CC BY 4.0 licensed, the TTS engine generates novel audio waveforms. The resulting synthetic speech is a new derivative work, not a direct reproduction of the original human voice.
- Source model: LibriTTS-R (CC BY 4.0)
Training Data Permissions
Generated audio files can be used freely for training machine learning models. There are no restrictions on using the synthetic speech for supervised learning, fine-tuning, or dataset augmentation. The input text is used only for synthesis and is not used for training the API's own models, ensuring clear data boundaries.
You retain full rights to the generated audio files. This makes the API suitable for building proprietary speech datasets where you need to control the data composition and licensing terms without external constraints.
Commercial Use Cases
The API is well-suited for commercial applications, including generating voiceovers for apps, educational content, or customer service bots. There are no monthly plans, seats, or minimum spend requirements.
The synthetic nature of the audio, combined with the clear data usage rights, makes it an ideal candidate for training datasets where legal clarity is paramount. You can distribute or sell datasets containing this synthetic speech without worrying about underlying voice licensing issues.
Performance Under Load
Processing requests sequentially means high concurrency introduces latency. Under heavy load, new synthesis jobs enter a queue, delaying output start times for individual requests.
Request Queue Behavior
Requests are processed one after another, not in parallel. During peak usage, a request may wait in the queue before synthesis begins, extending the total time to receive the audio file.
Generation Speed Estimates
Typical speed is roughly real time, generating about 25 characters of text per second of audio. Short phrases take a few seconds, while longer texts process at a consistent rate once dequeued.
Scalability for Bulk Jobs
For large ML datasets, split long texts into parts of up to 4,096 characters. The API returns separate audio files for each part, which your pipeline must join. This approach manages the single-threaded processing constraint effectively.

Language Support and Localization
Language coverage is a critical variable when scaling synthetic speech datasets. This API supports 23 distinct languages, allowing teams to generate diverse training data without managing multiple specialized services. The system accepts any of the eight voice presets for any supported language, ensuring consistent audio characteristics across multilingual corpora.
23 Languages Available
The service supports 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, providing flexibility in dataset construction.
Latin Script Defaults
If the language field is omitted, the system guesses the language from the script. Latin-script text defaults to English. For non-English Latin-script text, explicitly setting the language parameter is required to ensure accurate synthesis and prevent mispronunciation in your training data.
Unsupported Languages List
Vietnamese, Indonesian, Thai, Ukrainian, Romanian, Czech, Bengali, and all other languages not listed above are not supported. Ensure your target languages are included in the supported list before integrating this API into your data pipeline.
- Vietnamese
- Indonesian
- Thai
- Ukrainian
- Romanian
- Czech
Developer Experience and Integration
Integration is straightforward because the endpoint mirrors the OpenAI speech structure. You can reuse existing OpenAI SDKs by simply pointing the base_url to https://api.playlucky101.com/v1. Authentication requires a single Bearer key, which is issued upon account creation and can be rotated instantly if compromised.
The API returns standard HTTP status codes. Failed synthesis requests incur no charge, simplifying budget tracking for bulk generation jobs. This predictability allows ML teams to scale their dataset production without worrying about unexpected costs for failed audio generations.
Ease of API Setup
Setting up the API requires minimal configuration. By updating the base_url in your preferred OpenAI-compatible SDK, you gain immediate access to the TTS engine. The single active key per account reduces credential management overhead, and generating a new key instantly invalidates the old one for security.
Documentation Clarity
The documentation clearly defines the JSON request body, including required fields like model and input. It specifies the supported response formats (mp3, wav, opus, flac, pcm) and the character limit of 4096 per request. This clarity reduces integration time for developers building automated data pipelines.
Error Handling Mechanisms
The API distinguishes between successful synthesis and failures by returning appropriate HTTP status codes. Crucially, if synthesis fails, the charge is returned in full. This ensures that your cost-per-character metrics remain accurate even when processing large volumes of text with varying complexity.
Comparison with Other TTS Providers
We evaluate synthetic speech APIs for startups building ML datasets and speech products. Our reviews focus on cost-per-character, language coverage, and data rights for training pipelines.
Feature Set Differences
This API offers a standard OpenAI-compatible endpoint, simplifying integration for teams already using Python or Node.js SDKs. Unlike providers offering SSML or voice cloning, it provides eight fixed presets with no custom pronunciation control. The lack of streaming is acceptable for batch dataset generation but limits real-time applications.
- OpenAI-compatible
POSTendpoint - No SSML or emotion tags
- Fixed voice presets only
Pricing Competitiveness
At $10 per million characters, costs are predictable and transparent. Failed requests incur no charge, protecting budgets during noisy data pipelines. The USDT-only payment model avoids transaction fees associated with credit cards or PayPal, appealing to crypto-native startups.
- $0.01 per 1,000 characters
- Free failed requests
- USDT (TRC-20) payments only
Niche Advantages for Startups
The primary advantage is the licensing of source recordings under CC BY 4.0. This clear rights framework reduces legal risk when training commercial ML models. Combined with 23 language supports and no minimum spend, it is ideal for prototyping diverse speech datasets without enterprise contracts.
- CC BY 4.0 source data
- No minimum spend
- 23 languages supported
Limitations to Consider
Lucky TTS Review finds that while the core API is robust, certain constraints impact large-scale ML dataset workflows. Teams requiring precise phonetic control or branded voice identity must account for these specific limitations before integrating.
No Custom Voice Cloning
The service does not offer custom voice cloning, meaning you cannot upload a specific speaker's audio to create a unique voice model. If your dataset requires a proprietary voice identity, you must rely on the eight standard presets or use external cloning tools before feeding text to this API.
Limited Control Features
Unlike specialized TTS engines, there is no SSML support, phoneme dictionary, or custom pronunciation lexicon. You cannot fine-tune emotion tags or pause durations programmatically. This lack of granular control may require pre-processing text to ensure consistent pronunciation in your training data.
Single Key Constraint
Each account is limited to one active API key. If you lose or compromise your key, the old one stops immediately upon issuing a new one. For high-throughput ML pipelines, you must manage concurrency and error handling carefully, as a new key invalidates the old one instantly.
Final Recommendation for ML Teams
This API offers a cost-effective, compliant path for generating synthetic speech datasets at scale. With CC BY 4.0 source recordings and clear data rights, it suits startups prioritizing low-cost volume over granular voice control. The unified endpoint structure simplifies integration into existing ML pipelines without needing custom SDKs.
The pay-per-character model and USDT payments reduce financial friction, while the lack of streaming or SSML keeps the focus on bulk generation. Teams needing precise phoneme control or real-time inference should look elsewhere, but for static dataset creation, this solution balances speed, cost, and legal clarity effectively.
Best Use Cases
Ideal for startups building large-scale text-to-speech training sets where voice diversity and language coverage matter more than nuanced emotion. The 8 preset voices cover 23 languages, providing sufficient variation for model training without the overhead of managing multiple providers.
Suitable for projects requiring clear ownership of synthetic audio data, as inputs are not used for training and outputs carry inaudible watermarks. The simple pricing model allows precise budget forecasting for bulk character generation.
- Large-scale dataset generation
- Multi-lingual ML model training
- Budget-conscious startup pipelines
- Clear data ownership requirements
When to Choose This API
Choose this API when your primary constraints are cost-per-character and ease of integration. The standard OpenAI-compatible endpoint allows quick adoption using existing libraries, while the $0.50 trial lets you validate quality before committing to USDT top-ups.
It is the right choice if you do not need custom voice cloning, SSML, or real-time streaming. The lack of these features reduces complexity and cost, making it efficient for offline batch processing of thousands of audio files.
- Cost-sensitive bulk generation
- Standard API integration
- No real-time streaming needed
- Clear data usage rights
Alternatives for Specific Needs
If you require real-time streaming or WebSocket support, this API’s request-response model may feel limiting. Similarly, if precise SSML control or custom pronunciation lexicons are critical, you may need a provider with deeper TTS configuration options.
For projects needing voice cloning or emotion tags, this preset-based system lacks the necessary granularity. While it covers 23 languages, it does not support niche languages like Vietnamese or Thai, so those teams should seek specialized providers.
If you prefer credit card payments or need an SLA for enterprise uptime, this API’s USDT-only, no-SLA model may not fit. However, for most startups building ML datasets, the trade-off in flexibility is outweighed by the low cost and simplicity.
- Real-time streaming needs
- Custom voice cloning required
- Niche language support
- Enterprise SLA requirements
- Credit card payment preference
Listen to the 8 voices
Each sample below was generated by this API from the same sentence. Additionally, their licensing terms for synthetic data are less transparent than some alternatives.
23 languages
Any voice speaks any of these languages.
Sources and further reading
- Wikipedia
- Python documentation
- Node.js documentation
- OpenAI API reference: create speech
- MDN: Web Speech API
- openai-node on GitHub
- W3C WAI: making audio and video accessible
- TRON developer hub: TRC-20
- LibriTTS-R corpus (OpenSLR 141)
- Creative Commons Attribution 4.0
- openai-python on GitHub
- IETF RFC 7845: Ogg Opus
