Frequently asked questions
Is Hindi well-supported in TTS?
Yes, Hindi is one of our 23 supported languages. You can generate audio in Hindi using any of the eight voice presets. The system handles the synthesis directly, so you do not need external language packs. Just include Hindi text in your request body, and the API will produce the audio file ready for your dataset.
Do I need to prepare special text formats?
No special formats are required. You simply send plain text in the JSON body. The API accepts up to 4096 characters per request. If your text exceeds this limit, split it into smaller parts. There is no need for SSML or phoneme dictionaries, keeping your preprocessing pipeline simple and fast.
Can I generate thousands of hours of audio?
Yes, the API supports high-volume generation. Requests are processed sequentially, so long texts take roughly real-time to generate. You can split large texts into 4096-character chunks and join the resulting audio files. With a sufficient balance, you can scale your dataset size without worrying about infrastructure limits or complex scaling.
How do I ensure voice consistency?
Consistency is achieved by using the same voice preset for all requests. You have eight options, such as Aria or Felix. Since the engine is consistent, using the same voice ID ensures uniform tone and timbre across your entire dataset, which is critical for training reliable machine learning models.
What is the best format for ML datasets?
We support mp3, wav, opus, flac, and pcm. Uncompressed formats like wav or flac preserve audio quality, while mp3 or opus reduce storage costs. However, mp3 or opus can reduce storage costs. Choose the format that best fits your model’s input requirements and your storage budget.
Do I need to label the audio files?
The API returns the audio file but does not manage your labeling schema. You must handle metadata and labels in your own pipeline. The generated audio includes an inaudible watermark to mark it as synthetic, which you can use for dataset curation, but you define the semantic labels yourself.
Can I use the same voice for all datasets?
Absolutely. You can use any of the eight voices for any project. Whether you are building a Hindi dataset or an English one, you can reuse the same voice ID. This allows you to maintain a consistent brand voice across multiple different machine learning projects and datasets.
How do I handle different dialects?
We do not offer specific dialect presets. However, you can influence pronunciation by setting the language parameter to the appropriate ISO code. For non-English Latin-script text, explicitly setting the language ensures correct interpretation. The voices are versatile, but precise dialect control relies on your input text and language settings.
Is the API suitable for real-time inference?
No, this API is designed for batch dataset generation, not real-time inference. Requests are processed sequentially, so under load, a request may wait in a queue. Typical speed is roughly real-time for long texts, but short phrases may take a few seconds, making it unsuitable for low-latency applications.
Do I need a developer to use this?
A basic technical understanding helps, but it is not strictly required. You can use the provided curl examples or the OpenAI SDK. If you can send a JSON POST request with your API key, you can generate audio. Our documentation provides ready-to-use code snippets to get you started quickly.
How do I validate the audio quality?
You validate quality by listening to the generated files. Each file is returned directly in the API response. You can check for clarity, tone, and accuracy. The voices are based on licensed recordings, providing consistent quality. You can also verify the inaudible watermark to confirm the audio is synthetic.
Can I export the audio in batches?
Yes, you can generate audio in batches by sending multiple requests. Since the API processes requests sequentially, you can script your own batch jobs. Split your text into 4096-character chunks and send them one by one. The API returns each audio file, which you can then aggregate into your dataset.
What happens if the API goes down?
We do not provide an uptime guarantee or SLA. If the API is unavailable, your requests will fail. Failed requests are free, so you are not charged for audio that was not generated. You can retry your requests once the service is back online without any penalty or extra cost.
Do I need to pay for each voice?
No, you do not pay extra for different voices. The price is $10 per 1 million characters regardless of the voice or model used. Whether you use Aria, Felix, or any other preset, the cost per character remains the same. All eight voices are included in your standard rate.
Why Use Synthetic Hindi Speech?
Generating high-quality synthetic Hindi speech allows ML teams to overcome data scarcity without the logistical burden of managing human talent. Synthetic datasets provide the volume and variety needed to train robust models.
Data scarcity in Hindi
Hindi is a high-resource language compared to many others, but specific dialect variations and niche domains often lack sufficient training data. Synthetic generation fills these gaps, allowing you to create balanced datasets that accurately reflect real-world usage patterns without relying on limited human recordings.
- Access to consistent voice quality across unlimited text volumes
Cost of human recording
Traditional human recording involves significant overhead: hiring talent, booking studios, managing schedules, and editing audio. For ML datasets, where you might need thousands of hours of audio, these costs scale linearly. Synthetic speech eliminates these variable expenses, offering a predictable, flat rate of $0.01 per 1,000 characters regardless of voice talent fees.
- No studio rental or talent hourly rates
- Elimination of post-production editing costs
Scalability of API
Our API handles character limits up to 4096 per request, with automatic splitting for longer texts. This structure allows you to generate massive datasets programmatically. You can integrate this directly into your training pipelines, generating thousands of audio files in minutes rather than weeks, accelerating your model development cycle.
- Programmatic generation via standard HTTP requests
- Automatic handling of long text segments
Preparing Your Text Data
When building synthetic Hindi datasets, you must handle the 4,096-character limit per request. Text exceeding this threshold needs splitting before submission to avoid errors, then the resulting audio files are joined. This ensures consistent quality without hitting API boundaries.
Character limit management
Split long scripts into segments under 4,096 characters. Each segment generates one audio file. You then concatenate these files to reconstruct the full narration. This approach keeps requests small and manageable for your pipeline.
Text normalization
Clean your Hindi text before sending it to the API. Remove extra whitespace, standardize punctuation, and ensure consistent encoding. Normalized text reduces synthesis errors and produces cleaner audio files for your machine learning models.
- Strip extra spaces and line breaks.
- Standardize punctuation marks.
- Ensure UTF-8 encoding.
Language detection
The API guesses the language from the script if you do not specify it. Latin text defaults to English. For Hindi, the Devanagari script is recognized automatically, but explicitly setting the language parameter ensures accuracy. However, explicitly setting the language parameter ensures accuracy for mixed-script datasets.
Generating Audio Files
Sending text to our API is straightforward. You POST a JSON body to https://api.playzbzb.com/v1/audio/speech with your API key. The engine supports models like tts-1 and tts-1-hd, accepting up to 4096 characters per request. Ensure you specify the language field as 'hi' to guarantee correct Hindi pronunciation, as the system defaults to English for Latin script text.
API request structure
Include your text in the input field and select a voice preset like nova or felix. The response returns the audio file directly. For ML datasets, you can choose from MP3, WAV, OPUS, FLAC, or PCM formats depending on your training pipeline requirements.
- Endpoint: POST /v1/audio/speech
- Auth: Bearer tts-... key
- Max input: 4096 characters
- Languages: 23 supported, including Hindi
Voice selection
Select voices that match your dataset’s tone. Nova offers a bright, upbeat style, while Sage provides a calm, instructional voice. All eight voices support Hindi. The generated audio includes an inaudible watermark, marking it as synthetic speech for your model’s training data.
- Aria, Nova, Luna, Sage (Female)
- Felix, Milo, Atlas, Orion (Male)
- All voices speak all 23 languages
Format conversion
You can specify the output format directly in the API request using the response_format parameter. This eliminates the need for external conversion tools. Choose MP3 for compact storage or WAV/FLAC for lossless quality. PCM is available for raw audio processing.
- MP3, WAV, OPUS, FLAC, PCM
- Direct output via API parameter
- No external tools needed
Quality Control for ML
Ensuring data quality is critical when training machine learning models. Our synthetic audio files include an inaudible watermark, allowing you to distinguish generated speech from human recordings during dataset curation. This verification step helps maintain transparency and prevents contamination of your training sets.
Standardizing output formats ensures compatibility with your ML pipelines. By selecting specific audio formats like WAV or FLAC, you guarantee consistent bitrates and sampling rates, reducing preprocessing overhead and ensuring uniformity across your entire Hindi speech dataset.
Watermark verification
Every generated audio file carries an inaudible digital watermark marking it as synthetic speech. This allows your pipeline to automatically identify and tag synthetic data, ensuring clear distinction from human-recorded audio in your training sets.
You can verify the watermark presence programmatically or through audio analysis tools to confirm the synthetic nature of the data before model ingestion.
- Inaudible synthetic speech marker
- Automatable verification steps
- Clear data provenance
Audio format standards
Select your preferred response format (mp3, wav, opus, flac, pcm) to match your model’s input requirements. WAV and FLAC offer lossless quality ideal for training, while MP3 reduces storage costs.
Consistent formatting across all generated files minimizes preprocessing steps, ensuring your ML pipeline receives uniform data structures without unexpected codec variations.
- Lossless options: WAV, FLAC
- Compressed: MP3, OPUS
- Raw PCM for custom processing
Consistency checks
Maintain consistent voice and speed settings across your dataset to reduce variance in your training data. Using the same voice preset for related samples ensures tonal uniformity.
Regularly check character counts and request limits to avoid interruptions in your batch generation process, ensuring a steady flow of high-quality audio samples.
- Uniform voice presets
- Stable speed settings
- Regular limit monitoring
Integrating into Pipelines
Integrating TTS into your ML pipeline requires automating the request cycle to handle large synthetic speech datasets efficiently. By using standard HTTP clients, teams can generate audio files on demand without managing complex infrastructure or waiting for human narrators.
Automating these workflows ensures consistent data generation speeds, aligning with the API’s real-time processing capabilities while maintaining the structured input required for model training.
Automated workflows
Use the OpenAI SDK or plain HTTP to script requests. Set the base_url to https://api.playzbzb.com/v1 and pass your key. This allows seamless integration into Python or Node.js pipelines for continuous dataset generation.
- - Use standard HTTP clients for scripting
- Set
base_urltohttps://api.playzbzb.com/v1- Pass API key viaAuthorizationheader
Batch processing
Split long texts into 4096-character chunks and join the resulting audio files. Process requests sequentially to avoid rate limit issues, keeping in mind the 60 requests per minute limit per key.
- - Split text into 4096-character segments - Join resulting audio files - Process sequentially to respect limits
Error handling
Implement retry logic for failed requests, which are free. Monitor the response status to ensure each audio file is generated correctly before adding it to your training set. Check balance regularly to avoid interruption.
- - Implement retry logic for failures - Verify response status for each file - Monitor balance to prevent interruption
Voice Selection Strategies
Choosing the right voice ensures your synthetic dataset matches the intended use case. Hindi text works with any of our eight preset voices, allowing you to generate diverse training data without complex infrastructure. Consistency in voice selection helps maintain uniform quality across large ML datasets.
Pairing voice characteristics with content type improves model performance. Use neutral tones for news, energetic voices for promos, and soft narrators for storytelling to cover different linguistic patterns and tones in your training set.
Neutral tones for news
Atlas provides a neutral, news-style delivery ideal for journalistic datasets. His consistent pacing and clear articulation make him suitable for speech recognition models trained on formal Hindi content.
- Best for: News broadcasts, formal readings
- Voice: Atlas
Energetic tones for promo
Felix offers an energetic, promo-style voice that adds dynamism to your dataset. His upbeat tone helps models learn conversational and promotional Hindi speech patterns effectively.
- Best for: Advertisements, promotional content
- Voice: Felix
Storytelling voices
Luna and Aria provide soft, storytelling tones perfect for narrative datasets. Luna’s soft, storytelling tone suits audiobooks, while Aria’s warm, clear narration works well for engaging stories.
- Best for: Audiobooks, narratives
- Voices: Luna, Aria
Cost Optimization
Generating synthetic speech for ML datasets involves significant character volume, so understanding the exact cost structure helps you budget accurately. Our pricing is straightforward: $10 per million characters, which equals $0.01 per 1,000 characters. Spaces and punctuation count toward this total, so ensure your text normalization steps account for every character before submission.
Bulk character usage
To optimize costs for large datasets, split texts into chunks of up to 4096 characters. This ensures efficient processing and allows you to join audio files seamlessly. Since requests are processed sequentially, larger batches may take longer due to queueing.
- Calculate total characters including spaces and punctuation.
- Split input text into 4096-character segments.
- Join output audio files programmatically.
Failed request savings
If synthesis fails, the charge is returned in full. This protects your budget during dataset creation, especially when testing different voices or formats. You only pay for successful audio generations, making it risk-free to experiment with various voice settings for your ML models.
USDT efficiency
Top-ups are handled via USDT on the TRON network (TRC-20), with a minimum of $5. Credits are added automatically within minutes, ensuring your team never stops generating audio due to payment delays. There are no monthly plans or hidden fees, so your balance does not expire.
- Minimum top-up: $5.
- Network: TRON (TRC-20).
- No expiration on balance.

Language Support Details
Supporting Hindi is straightforward because our engine handles the script natively. You simply provide the Devanagari text and set the language parameter to ensure accurate pronunciation. The API processes up to 4096 characters per request, allowing you to feed substantial corpus segments directly without complex pre-segmentation.
Hindi script support
Send standard Devanagari Unicode in the input field. Our model recognizes the characters and applies correct phonetic rules. No special preprocessing is needed beyond standard text normalization.
- Accepts standard Unicode Devanagari text
- No custom lexicons required
Other language options
Hindi is one of 23 supported languages, including English, Spanish, and Arabic. You can generate mixed-language datasets or switch voices and languages per request. This flexibility helps create diverse training data for multilingual models without managing multiple APIs.
- 23 total languages supported
- Mix languages across requests
Fallback behavior
If you omit the language field, the API guesses based on the script. Latin text defaults to English, while Hindi Devanagari is recognized automatically. However, explicitly setting language ensures consistent results, especially for mixed scripts or less common language variants.
Common Pitfalls
When generating Hindi TTS at scale, overlooking rate limits can stall your pipeline. The API allows 60 requests per minute per key, dropping to 10 before the first top-up. Queueing delays are normal under load, so build retry logic with exponential backoff to handle temporary bottlenecks without wasting characters.
Another frequent error is omitting the language field for Latin-script Hindi text, which defaults to English pronunciation. Always specify language: hi to ensure accurate phoneme generation. Additionally, exceeding the 4096-character limit per request causes failures; split longer texts into manageable chunks before sending.
Ignoring rate limits
The API enforces a strict limit of 60 requests per minute per key. If you exceed this, requests may queue or fail. This is critical for large ML datasets where thousands of sentences are generated quickly.
Use exponential backoff in your automation scripts. This prevents overwhelming the API during peak generation times and ensures you do not burn credits on failed attempts due to throttling.
- Rate limit: 60 requests/minute per key
- Pre-top-up limit: 10 requests/minute
- Queued requests may experience delays
Missing language codes
If the language parameter is not set, the system guesses the language based on the script. For Hindi written in Latin script (transliteration), it defaults to English, resulting in incorrect pronunciation.
Always include "language": "hi" in your JSON body for Hindi text to ensure the correct voice model applies the proper phonemes and intonation patterns specific to the language.
Large payload issues
Each API request accepts a maximum of 4096 characters in the input field. Sending longer texts in a single request will result in a failure and potential charge if not handled correctly.
Split your source text into segments under 4096 characters. Generate audio for each segment individually and concatenate the resulting audio files in your pipeline to maintain the full narrative flow.
Best Practices for Datasets
For synthetic speech datasets, consistency is key. Use multiple voices to avoid model bias, but keep generation parameters like speed and format uniform across your corpus to ensure reliable training data.
Diverse voice usage
Leverage all eight available voices to capture varied speech patterns. Rotate between male and female presets like Aria and Felix to create a balanced dataset that improves model robustness.
Consistent formatting
Standardize your input text format before sending requests. Ensure all text stays within the 4096-character limit and use consistent encoding to prevent silent truncation or encoding errors during synthesis.
Metadata tagging
Tag each audio file with its source voice, language code, and speed setting. This metadata helps you filter and select specific segments for training without re-generating the audio.
- Record voice ID and generation timestamp for audit trails.
- Tag with ISO language code even if auto-detected.
- Include character count used for accurate cost tracking.
Troubleshooting
Troubleshooting synthetic speech generation requires checking specific API responses and account states. Most issues stem from payload constraints or connectivity timeouts rather than engine failures.
Address these common hurdles quickly to keep your ML dataset pipeline moving without unnecessary delays.
Audio generation errors
If synthesis fails, no charge is applied, and the error details indicate whether the input exceeded the 4096 character limit or if the voice parameter was invalid.
- Check input length does not exceed 4096 characters.
- Verify voice names match the 8 presets exactly.
API connectivity issues
Requests may queue under load, causing delays. Ensure your client handles HTTP status codes correctly and retries on temporary failures.
- Monitor rate limits (60 req/min).
Balance problems
Check your balance in the dashboard. Failed requests are free, but insufficient funds halt generation. Top-ups via USDT on TRON take minutes to credit.
Future Scaling
As your synthetic speech needs grow, the API handles scale seamlessly. Generate thousands of hours of Hindi audio by splitting large texts into 4096-character chunks and joining the results. The system processes requests sequentially, so under heavy load, generation may queue but remains reliable for bulk dataset creation without complex infrastructure.
Increasing volume
Long texts are split automatically into parts of up to 4096 characters. The API returns separate audio files for each part, which you join into a single continuous file. This ensures consistent quality even for multi-hour corpora used in training large language models.
- Split input at 4096 characters per request
- Join output audio files sequentially
- Handle queue delays during high load
Adding languages
The same API endpoint supports 22 other languages, allowing you to expand your ML datasets without changing infrastructure. Each voice can speak all supported languages, ensuring consistent tone across multilingual models. Just update the language parameter in your requests to switch between Hindi, English, Spanish, or any of the 23 supported languages.
- Support for 23 languages in one API
- Voice consistency across language switches
- No new integration required for expansion
Team coordination
Each account is tied to a single active API key, which you can rotate immediately if compromised. Teams can coordinate by distributing the key. Since there are no shared seats, your team shares the same base URL structure and rate limits, simplifying access management for large groups generating datasets.
- Single active API key per account
- Immediate key rotation if compromised
- Distributed key usage for team members
Listen to the 8 voices
Each sample below was generated by this API from the same sentence. This ensures data provenance and prevents accidental mixing of human and synthetic samples during model training.
23 languages
Any voice speaks any of these languages.
