Frequently asked questions
Which voice is best for narration?
Aria is designed as a warm, clear narrator, making it highly suitable for general narration. Luna is also a strong option, particularly for storytelling contexts. Both voices deliver consistent, high-quality synthetic audio across all 23 supported languages, ensuring your ML datasets maintain a professional and coherent tone regardless of the specific language or use case.
Can I use the same voice for multiple languages?
Yes, every voice can speak any of the 23 supported languages. You simply set the language parameter in your API request to the desired ISO code. This allows you to generate consistent synthetic speech datasets across multiple languages without needing to manage different voice models.
Do all voices support all 23 languages?
Yes, all eight voices support all 23 languages. You can generate audio in Arabic, Chinese, English, or any other supported language using any voice preset. This flexibility ensures you can build diverse, multilingual datasets for your ML models without being restricted by voice-language pairings.
What is the difference between Aria and Luna?
Aria is characterized as a warm, clear narrator, while Luna is described as soft and suited for storytelling. Both are female voices available for dataset generation. Choose Aria for clear, direct narration and Luna for softer, narrative-driven content depending on your specific dataset requirements.
Is Felix too energetic for datasets?
Felix is described as energetic and promotional. If your dataset requires high-energy audio for marketing or promo models, Felix is suitable. For more neutral or calm datasets, consider Atlas or Sage. You can adjust the speed (0.25–4.0) to fine-tune the perceived energy level.
Can I change the speed of the voice?
Yes, you can adjust the speed between 0.25 and 4.0 using the speed parameter. This allows you to generate audio at different rates for your datasets, which can be useful for training models on varied speech patterns or optimizing inference time for specific use cases.
Do voices sound different in different languages?
The voices remain consistent across languages, but pronunciation and intonation adapt to the specific language. Since voices are based on licensed recordings, they maintain their core character. You can specify the language parameter to ensure accurate generation for each of the 23 supported languages.
What is the best voice for news datasets?
Atlas is described as neutral and suited for news, making it a strong choice for news datasets. Milo is friendly and conversational, which may also work for spoken news content. Both provide consistent quality, allowing you to train models on realistic speech patterns for news applications.
Are the voices based on real people?
Yes, the voices are based on licensed recordings from LibriTTS-R (CC BY 4.0), derived from LibriVox public-domain audiobook readings. Each generated file includes an inaudible watermark marking it as synthetic speech, ensuring clarity about the audio's origin for your machine learning datasets.
Can I clone a voice for my dataset?
No, custom voice cloning is not offered. You cannot upload your own voice or design custom voices. You must choose from the eight available presets.
How do I choose between male and female voices?
Choose based on your dataset's target demographic or use case. Female voices include Aria, Nova, Luna, and Sage. Male voices include Felix, Milo, Atlas, and Orion. You can mix voices in your dataset by making separate API requests for each voice, allowing for diverse training data.
Is Nova good for marketing datasets?
Nova is described as bright and upbeat, making it well-suited for marketing datasets. Its energetic tone aligns well with promotional content. You can also adjust the speed to tailor the audio to your specific requirements.
Does Orion sound too deep for general use?
Orion is described as deep and suited for documentaries. While deep, it is a standard voice preset available for general use. If the depth is too pronounced for your specific dataset, you can adjust the speed or choose a more neutral voice like Atlas.
Can I mix voices in one dataset?
Yes, you can generate audio using different voices by making separate API requests. Each request produces an audio file that you can combine into your dataset. This allows you to create diverse training data with varied vocal characteristics from the eight available presets.
Female voices overview
Selecting the right female voice is critical for building consistent synthetic speech datasets. Our API offers three distinct presets designed to cover the most common narrative styles found in training data for voice models.
Aria: warm narrator
Aria provides a clear, warm tone ideal for general narration and audiobook-style datasets. Its balanced delivery ensures high intelligibility across diverse text samples, making it a reliable baseline for model training.
Nova: bright upbeat
Nova delivers a bright, energetic sound perfect for marketing content or youth-oriented datasets. This voice adds dynamism to synthetic data, helping models learn varied prosody and emotional range.
Luna: soft storytelling
Luna offers a softer, more intimate tone suited for storytelling or calm instructional content. Its gentle cadence helps create datasets that mimic natural, relaxed speech patterns.
Male voices overview
Male voices offer distinct tonal ranges essential for diverse synthetic speech datasets. Selecting the right male preset depends on whether your model requires high-energy promotional material, natural conversational flow, or authoritative news delivery.
VoxyData provides four specialized male voices, each optimized for specific dataset characteristics to ensure consistent audio quality across your ML training data.
Felix: energetic promo
Felix delivers an energetic tone ideal for promotional content and advertising datasets. His voice carries a dynamic quality that captures attention, making it suitable for short-form marketing audio.
Use Felix when your model needs to learn from high-energy, upbeat commercial speech patterns.
- Ideal for promotional and advertising datasets
- Dynamic energy for short-form content
Milo: friendly conversational
Milo provides a friendly, conversational tone that mimics natural human dialogue. This voice is perfect for chatbot training data or interactive voice response (IVR) systems.
His approachable delivery helps models learn casual, everyday speech patterns with high clarity.
- Best for conversational and chatbot datasets
- Approachable, natural dialogue style
Atlas: neutral news
Atlas offers a neutral, clear delivery suitable for news reading and informational content. His voice remains consistent and professional, avoiding emotional bias.
Use Atlas for datasets requiring objective, clear speech without distinct personality markers.
- Neutral tone for news and informational data
- Consistent, professional delivery
Voice selection criteria
Selecting the right voice is critical for building high-quality synthetic speech datasets. The choice depends on the dataset style, target audience, and specific model requirements.
Dataset style
Match the voice to the intended use case. Narrative styles benefit from clear, steady tones like Aria or Luna, while dynamic content suits Felix or Nova.
- Narrative: Aria, Luna
- Dynamic: Felix, Nova
- Conversational: Milo
- News: Atlas
Target audience
Consider who will interact with the audio. Friendly tones like Milo work well for customer support simulations, while neutral voices like Atlas suit news aggregators.
Model requirements
Ensure consistency across your dataset. Use a single voice for uniform quality, or vary voices to train robust models. Keep text under 4096 characters per request for reliable generation.
- Consistency: Single voice
- Variety: Multiple voices
- Limit: 4096 chars
Aria vs. Luna
Aria and Luna offer distinct narrative styles for synthetic speech datasets. Aria provides a warm, clear narration ideal for structured content, while Luna delivers a soft, storytelling tone suited for immersive audio experiences.
Narrative clarity
Aria’s clear enunciation ensures high intelligibility, making it excellent for training models on structured text. Luna’s softer delivery adds emotional depth, which is valuable for narrative datasets requiring nuanced tone representation.
- Aria excels in clear, direct narration
- Luna adds soft storytelling nuances
Softness level
Luna’s voice has a gentler quality, reducing harshness in generated audio. Aria maintains a consistent, professional warmth without sacrificing clarity, ensuring the voice remains engaging without being overly dramatic.
- Luna offers a gentler, softer tone
- Aria maintains consistent professional warmth
Use case fit
Choose Aria for educational content, audiobooks, or clear instructional datasets. Select Luna for fiction, emotional storytelling, or datasets requiring a more intimate and engaging listening experience.
- Aria: educational content and audiobooks
- Luna: fiction and emotional storytelling
Felix vs. Milo
Choosing the right voice depends on whether you need high-impact promo material or natural conversational data. Felix and Milo offer distinct profiles for synthetic speech datasets.
Felix brings energetic, promo-style delivery ideal for marketing samples. Milo provides a friendly, conversational tone suitable for chatbot training data. Both voices maintain consistent quality across different text inputs.
Energy levels
Felix delivers higher energy with a dynamic pace, making it suitable for promotional content and high-engagement scenarios.
Milo maintains a moderate, steady energy level that feels natural and approachable, ideal for conversational datasets.
- Felix: energetic, promo-focused
- Milo: friendly, conversational
Conversational tone
Milo’s tone is designed to mimic natural dialogue, making it perfect for training language models on casual interactions.
Felix’s tone is more structured and clear, better suited for structured audio data like advertisements or announcements.
Promo vs. chat
Use Felix for datasets requiring clear, persuasive speech patterns typical in marketing or advertising.
Choose Milo for datasets capturing everyday conversation, customer service interactions, or casual dialogue scenarios.
Atlas and Orion
Atlas and Orion provide distinct tonal anchors for structured audio datasets. Atlas delivers a neutral, consistent delivery ideal for news reading and factual content. Its voice remains steady and professional, ensuring that synthetic speech sounds authoritative without emotional bias. This consistency is critical when training models on news or informational text where clarity outweighs personality.
Orion offers deeper, documentary-style narration. It suits longer-form scripts where a grounded, serious tone is required. The voice carries weight and gravity, making it well-suited for educational or historical datasets. Using Orion ensures your model learns from audio that feels authoritative and immersive, rather than casual or promotional.
Neutral delivery
Atlas is designed for clarity and neutrality. It avoids strong emotional inflections, making it perfect for news reading tasks. The voice remains consistent across various topics, ensuring your dataset maintains a uniform professional tone. This stability helps ML models learn accurate pronunciation and pacing without being distracted by stylistic variations.
- Ideal for news and factual content
- Consistent professional tone
- Low emotional variance
Documentary depth
Orion brings a deeper, more resonant quality to synthetic speech. It is optimized for documentary-style narration, providing a sense of gravity and seriousness. This voice works well for educational materials, historical records, or long-form storytelling where a compelling, grounded voice is needed to maintain listener engagement over extended periods.
- Deep, resonant tone
- Suitable for long-form scripts
- Enhances narrative immersion
News reading
For datasets focused on current events or information dissemination, Atlas is the optimal choice. Its neutral delivery ensures that the content remains the focus. The voice is clear and articulate, making it easy for speech recognition models to process. This makes Atlas particularly valuable for training models on structured, text-heavy audio data.
Orion can also be used for news if a more dramatic or investigative tone is desired. However, for standard news reading, Atlas provides the reliable, unbiased baseline that most news datasets require. Both voices support all 23 languages, allowing for diverse multilingual news datasets.
- Atlas for standard news
- Orion for investigative tone
- Supports all 23 languages
Nova for upbeat content
Nova brings a bright, upbeat quality to synthetic speech datasets, making it ideal for high-energy content. Its consistent tone ensures uniformity across large-scale dataset generation, which is critical for training robust voice models.
The voice handles varied scripts well, maintaining clarity even with rapid pacing. This makes Nova a reliable choice for product teams needing scalable audio data without sacrificing auditory appeal.
High energy
Nova delivers consistent energy levels across different text lengths and complexities. This stability is crucial for ML datasets where variation in delivery can introduce noise into model training.
The voice maintains its upbeat character even under speed adjustments, allowing engineers to simulate diverse speaking rates without losing the core tonal identity essential for accurate classification.
- Consistent energy across varying text lengths
- Stable performance with speed adjustments
- Ideal for high-paced audio samples
Youthful tone
The youthful quality of Nova appeals to younger demographics, making it suitable for datasets targeting specific audience segments. This tone remains clear and distinct.
- Appeals to younger demographic segments
- Clear timbre for noisy environments
- Distinct from deeper voice profiles
Marketing data
Nova is particularly effective for marketing-focused datasets, where enthusiasm drives engagement. Its bright tone mirrors the energy often found in commercial audio content, ensuring realistic training data.
Use Nova to generate promo-style audio at scale. With character-based pricing, you can generate large volumes of marketing samples efficiently, keeping costs predictable while maintaining high quality.
- Ideal for promo-style audio
- Bright tone mirrors commercial content
- Cost-effective for large marketing datasets

Language compatibility
VoxyData’s TTS API supports 23 languages, including English, Spanish, German, Japanese, and Hindi. This range covers most major markets for synthetic speech datasets, ensuring broad applicability for training models that need diverse linguistic inputs.
23 languages supported
The API handles Arabic, Chinese, Danish, Dutch, Finnish, French, Greek, Hebrew, Italian, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Swahili, Swedish, and Turkish. If a language isn’t listed, it’s not supported, so verify your target languages before generating large batches.
- Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Turkish
Voice language agnostic
Any of the 8 voices can speak any of the 23 languages. This decoupling allows you to maintain consistent voice identity across multilingual datasets, which is critical for training robust voice models that generalize well across languages.
Script detection
If you don’t specify a language, the API guesses it from the script. Latin text defaults to English, which can cause errors for non-English Latin-script text like German or Spanish. Always set the language parameter explicitly for non-English inputs to ensure accurate synthesis.
Technical considerations
Technical parameters directly impact dataset utility. Selecting the right format and speed ensures your synthetic audio aligns with model training requirements without unnecessary overhead.
Speed and character limits dictate generation efficiency. Understanding these constraints helps you batch-process large volumes consistently while maintaining audio quality.
Audio format impact
['Choose mp3 for compact storage or wav for lossless fidelity. opus and flac offer balanced compression, while pcm provides raw data for specific preprocessing pipelines. Your choice affects file size and downstream processing speed.', 'Speed adjustments (0.25–4.0) allow you to simulate varied speaking rates. Generation time scales with speed; short phrases take a few seconds, and longer texts generate at roughly real time.'] ['mp3: Small size, broad compatibility', 'wav: Lossless, ideal for training', 'opus: Efficient compression', 'flac: Lossless compression', 'pcm: Raw audio data']
Character limits
['Each request accepts up to 4096 characters. For longer texts, split the input into chunks and join the resulting audio files. This ensures consistent quality and avoids truncation.', 'Spaces and punctuation count toward the limit. Plan your text segmentation carefully to maintain narrative flow across generated segments. Failed requests incur no charge, but successful ones consume your balance at $0.01 per 1,000 characters.'] ['Max 4096 characters per request', 'Split and join for longer texts', 'Spaces/punctuation count as characters']
Dataset consistency
When building synthetic speech datasets, consistency ensures your model learns reliable patterns rather than artifacts. Using the same voice preset across thousands of text samples guarantees that prosody and timbre remain uniform, which is critical for training robust voice models.
Same voice, different text
Keep the voice preset constant while varying the input text. This isolates linguistic variables from speaker-specific traits, allowing your ML model to focus on pronunciation and intonation patterns rather than adapting to different timbres.
Uniform quality
VoxyData’s API delivers consistent audio quality across requests. Since each generation uses the same underlying engine parameters, the resulting audio files maintain similar noise floors and clarity, reducing the need for extensive post-processing cleanup in your dataset.
Batch generation
Generate large volumes of audio by splitting texts into 4096-character chunks. Process these sequentially or in parallel to build comprehensive datasets. The consistent character-based pricing ($0.01 per 1,000 characters) helps you budget for massive data collection efficiently.
Watermarking synthetic data
Every audio file generated by VoxyData contains an inaudible watermark that identifies it as synthetic speech. This feature is essential for maintaining transparency in ML datasets, ensuring that downstream models can distinguish between human-recorded and AI-generated training data.
The watermark is embedded directly into the audio waveform without affecting perceptual quality, making it seamless for voice model training while preserving dataset integrity for researchers who need to audit their data sources.
Inaudible watermarks
The synthetic speech watermark is embedded during synthesis and persists through standard format conversions. It remains inaudible to the human ear, ensuring that your training data remains auditable without compromising audio fidelity for your models.
- Embedded during synthesis
- Persists across formats
- Inaudible to human listeners
Identification
You can verify synthetic audio using standard detection tools compatible with the embedded signal. This allows you to audit your dataset composition accurately, confirming the ratio of synthetic to human voice data before model training begins.
Verification is straightforward because the watermark is present in all generated files.
Dataset labeling
Include metadata tags in your dataset schema to mark files as synthetic. This helps downstream models learn to handle synthetic inputs appropriately and prevents contamination when mixing synthetic data with human-recorded audio.
Labeling is simplified because every generated file carries the same watermark signature, allowing you to programmatically identify and categorize synthetic audio at scale without manual review of individual files.
Summary of recommendations
Choosing the right voice ensures your synthetic dataset aligns with specific use cases. Consistency in tone and style is critical for training models that perform well in production environments.
Use the recommendations below to match voice characteristics with your target application, ensuring high-quality, uniform audio for your ML pipelines.
Best for narration
Aria offers a warm, clear narrator style ideal for audiobooks and long-form content. Luna provides a softer, storytelling tone suitable for immersive narratives.
- Aria: warm, clear narrator
- Luna: soft, storytelling
Best for conversation
Milo delivers a friendly, conversational tone perfect for chatbots and interactive assistants. Nova adds a bright, upbeat energy for engaging user interfaces.
- Milo: friendly, conversational
- Nova: bright, upbeat
Best for news
Atlas provides a neutral delivery suitable for news reading and factual content. Orion offers deep, documentary-style voice for serious, authoritative material.
- Atlas: neutral, news
- Orion: deep, documentary
Listen to the 8 voices
Each sample below was generated by this API from the same sentence. This helps prevent overfitting and improves model generalization.
23 languages
Any voice speaks any of these languages.
