Frequently asked questions
Which voice is best for audiobooks?
Q: Which voice is suitable for audiobooks? A: Luna is described as soft and storytelling, making it suitable for audiobooks. Aria is also a warm, clear narrator option. Since voices are based on licensed recordings from LibriVox, they are designed for long-form content. You can select either voice for your text-to-speech API requests to generate audio files.
Can male voices speak French well?
Yes. Any of the eight voices can speak any of the 23 supported languages, including French. You do not need a specific male voice for French; all presets like Felix, Milo, Atlas, and Orion support French. Simply set the language parameter to 'fr' in your API request.
What is the difference between Aria and Luna?
Aria is characterized as a warm, clear narrator, while Luna is described as soft and storytelling. Both are female voices available in the catalog. You choose between them based on whether you prefer a clear narrative style or a softer, more story-like tone for your text-to-speech output.
Is Orion suitable for news broadcasts?
Atlas is the primary choice for news broadcasts, described as neutral and news-oriented. Orion is deep and documentary-style. While you can use Orion, Atlas is specifically designed for a neutral news tone. Select the voice that best matches your specific broadcast style requirements.
Do voices sound different in Hindi?
The voices remain the same presets regardless of language. However, the audio output will reflect the phonetics and intonation of Hindi. You must set the language parameter to 'hi' for accurate pronunciation. The voice characteristics like warmth or depth remain consistent across all 23 supported languages.
Can I change a voice's speed mid-sentence?
No. The speed parameter applies to the entire request, ranging from 0.25 to 4.0. You cannot change speed dynamically within a single audio file. To vary speed, you must split the text into separate requests, generate individual audio files, and join them programmatically.
Are the voices gender-neutral in some languages?
No. The eight presets are distinctly categorized as female (Aria, Nova, Luna, Sage) or male (Felix, Milo, Atlas, Orion). There are no gender-neutral voice options listed. You must select a specific preset based on the desired gender tone for your application's audio output.
How do I know if a voice is licensed?
Every generated file carries an inaudible watermark that marks it as synthetic speech. The voices are based on licensed recordings from LibriTTS-R (CC BY 4.0) and LibriVox public-domain audiobook readings. This licensing information is part of the product facts and applies to all eight presets.
Can I use multiple voices in one project?
Yes. You can make separate API requests for each segment of text using different voice presets. Since the API returns audio files, you can join these files together in your application. Each request is charged separately based on the character count of the input text.
What is the best voice for kids?
Q: What voice is suitable for kids? A: There is no specific voice designated for children. You can use any of the eight presets, such as Nova for a bright, upbeat tone or Luna for a soft, storytelling style. The choice depends on your creative preference. All voices support the same 23 languages and formats.
Does Nova sound too energetic for drama?
Nova is described as bright and upbeat. If this energy level does not fit your drama, you can choose Luna, which is soft and storytelling, or Aria, which is a warm narrator. You can also adjust the speed parameter between 0.25 and 4.0 to modify the delivery pace.
Can Atlas be used for podcasts?
Atlas is described as neutral and news-oriented, making it suitable for podcasts requiring a clear, professional tone. Milo is friendly and conversational, which may also work well. You can use any voice preset for podcasts by setting the appropriate voice parameter in your API request.
Are there any voice-specific rate limits?
No. Rate limits apply per API key, not per voice. You have a limit of 60 requests per minute per key (10 per minute before the first top-up). This limit is consistent regardless of which of the eight voices you select for your text-to-speech requests.
How do I test a voice before buying?
Sign in with Google to receive a free trial of $0.50 credit, valid for 7 days. This allows you to generate approximately 50,000 characters of audio. You can test all eight voices and formats. Email/password accounts start without trial credit and require a top-up to begin testing.
Female Voice Options
Lumivoice provides four distinct female voice presets, each derived from licensed LibriTTS-R recordings. These voices are designed to offer consistent, scalable audio generation for indie developers who need specific tonal qualities without complex infrastructure.
Selecting the correct voice depends on your application’s genre. Each preset carries specific acoustic characteristics that influence user perception, from warm narration to bright upbeat delivery.
Aria: Warm and Clear Narrator
Aria delivers a warm, clear tone suitable for general narration. Her voice is designed for clarity, making her ideal for applications requiring straightforward text-to-speech output without emotional complexity.
Nova: Bright and Upbeat
Nova offers a bright, upbeat delivery style. This voice is optimized for content that requires energy and positivity, such as casual apps or interactive guides.
Luna: Soft Storytelling Tone
Luna provides a soft, storytelling tone. Her delivery is gentle and engaging, making her well-suited for narrative-driven applications or longer-form content consumption.
Sage: Calm and Instructional
Sage is characterized by a calm, instructional demeanor. She is designed for educational tools or guided experiences where a steady, reassuring voice aids user comprehension.
Male Voice Options
The male voice catalog offers four distinct presets tailored to specific project tones. Each voice is derived from licensed recordings, ensuring consistent quality across your application’s audio output without requiring complex infrastructure management.
Felix: Energetic Promotional Style
Felix delivers an energetic delivery ideal for promotional content and short marketing clips. His tone maintains high engagement levels suitable for time-sensitive announcements.
Milo: Friendly and Conversational
Milo provides a friendly, conversational sound that works well for user interfaces and assistant-like interactions. He creates a natural, approachable listening experience for everyday tasks.
Atlas: Neutral News Anchor
Atlas offers a neutral, news-style delivery appropriate for informational content. His consistent pacing ensures clarity for factual reporting and data-driven audio outputs.
Orion: Deep Documentary Style
Orion features a deep, documentary-style voice suited for narrative content. His lower register provides authority and depth for long-form storytelling or explanatory videos.
Matching Voices to Genres
Selecting the appropriate voice model ensures your application’s audio aligns with content expectations. Lumivoice offers distinct presets optimized for specific narrative tones, allowing indie developers to maintain consistent brand identity across diverse media types.
Fiction and Storytelling
Luna provides a soft, storytelling tone ideal for novels or audiobooks. Aria serves as a warm, clear narrator for general fiction. These voices prioritize emotional resonance over technical precision.
Educational and Instructional
Sage delivers a calm, instructional voice suitable for tutorials and e-learning modules. Its neutral delivery supports clarity without distracting from the core educational material.
Promotional and Marketing
Felix offers an energetic, promotional style perfect for advertisements or app announcements. Nova provides a bright, upbeat tone that engages listeners in commercial contexts.
News and Reporting
Atlas presents a neutral, news-anchor style for factual reporting. Orion’s deep, documentary style suits long-form investigative content or serious narrative reports.
Language Considerations
Lumivoice supports 23 languages, allowing any of the eight preset voices to speak any supported language. This flexibility simplifies integration for indie developers building multilingual apps, as you do not need to manage separate voice models for different linguistic regions. The system ensures consistent audio quality across languages like Arabic, Japanese, or Swahili without requiring additional configuration beyond the standard API call.
Voice Performance in Different Languages
Each of the eight voices delivers clear speech across all 23 supported languages, including complex scripts like Chinese, Korean, and Arabic. There is no need to assign specific voices to specific languages; for example, Aria can narrate in French with the same clarity as in English. This uniformity reduces decision fatigue during development and ensures predictable audio output regardless of the target language.
- Supports 23 languages including English, Spanish, Hindi, and German.
- Any voice preset works with any supported language.
- No separate voice models required per language.
Setting Language Codes Explicitly
While the API can guess the language from the script, explicit language codes ensure accurate pronunciation and intonation. For non-English texts using Latin script, such as Dutch or Malay, you must set the language field. This prevents the system from defaulting to English, which would result in incorrect accentuation and reduced comprehension for end-users.
- Always set the
languageparameter for non-English Latin text. - Use ISO 639-1 codes (e.g.,
de,fr,ja). - Explicit codes improve pronunciation accuracy.
Handling Latin Script Ambiguity
If the language field is omitted, Latin-script text defaults to English. This default behavior can cause issues for words like 'cafe' or 'naive' which exist in multiple languages but may be pronounced differently. Always specify the language code to avoid mispronunciation. Supported languages include Vietnamese and Indonesian. Thai, Ukrainian, and Romanian are not supported.
- Latin text defaults to English if
languageis unset. - Specify codes for all non-English Latin-script inputs.
- Not all Latin-script languages are supported.
Technical Voice Properties
Technical parameters define how the generated audio behaves in your application environment. Understanding speed ranges, format choices, and synthetic markers ensures consistent playback and correct attribution of the generated content.
Speed Adjustment Range
The API accepts a speed value between 0.25 and 4.0. This allows precise control over playback duration without altering voice quality. Faster speeds reduce generation time for long texts, while slower rates improve clarity for complex instructions.
- Minimum speed: 0.25x
- Maximum speed: 4.0x
- Default speed: 1.0x
Audio Format Impact
Select mp3, wav, opus, flac, or pcm via response_format. MP3 offers the smallest file size for web delivery, while WAV and FLAC provide lossless quality. PCM is raw data, useful for direct hardware integration. Format choice directly affects bandwidth and storage costs.
Watermark Identification
Every generated audio file contains an inaudible watermark identifying it as synthetic speech. This marker remains embedded regardless of the selected format or speed setting. It allows downstream systems to verify that the audio was generated by the Lumivoice API and not recorded by humans.
Licensing and Rights
Lumivoice voices are derived from licensed recordings in the LibriTTS-R dataset (CC BY 4.0) and public-domain readings from LibriVox. This foundation ensures that our audio generation relies on verified, accessible source material rather than proprietary black-box models.
The CC BY 4.0 license permits commercial use with attribution, providing indie developers with a clear legal baseline. While our specific synthetic outputs carry distinct markers, the underlying voice data remains transparently sourced, reducing ambiguity for apps distributing audio globally.
LibriTTS-R Source Material
Our eight voice presets are built upon the LibriTTS-R dataset, which aggregates high-quality speech from LibriVox public-domain audiobooks. This ensures that the acoustic characteristics of voices like Aria, Nova, and Atlas are grounded in real, licensed human recordings rather than purely synthetic neural approximations.
- Source: LibriTTS-R dataset
- Origin: LibriVox public-domain readings
- Quality: High-fidelity speech samples
CC BY 4.0 Implications
The Creative Commons Attribution 4.0 International license governs the underlying voice data. This allows for broad commercial integration within your indie application without requiring complex royalty negotiations. You may distribute the generated audio freely, provided you respect the general terms of the dataset license.
- Commercial use permitted
- Attribution required per license terms
- No royalty payments due
Synthetic Speech Markers
Every audio file generated via the API includes an inaudible watermark. This digital signature identifies the output as synthetic speech, distinguishing it from human-recorded audio. This marker aids in content verification and metadata tracking for your end-users without affecting audio quality or playback.
- Inaudible digital watermark
- Identifies synthetic origin
- No impact on audio fidelity
Testing and Evaluation
To validate voice suitability, generate synthetic audio using representative text samples. Submit requests via the API endpoint and review the output files for tonal accuracy. This method confirms whether the selected preset aligns with your application’s narrative tone before scaling integration.
Listening to Sample Texts
Use the provided text input limits (up to 4096 characters) to create test payloads. Download the resulting audio files and play them in your target environment. This step verifies that the voice clarity matches expectations for your specific content type.
- Submit short phrases to check initial tone.
- Test full-length stories for narrative consistency.
- Verify pronunciation accuracy in target languages.
Comparing Voice Speeds
Adjust the speed parameter between 0.25 and 4.0 to evaluate pacing. Faster rates may reduce clarity, while slower rates enhance instructional tone. Generate multiple versions to identify the optimal balance for your audience's listening habits.
- Test 1.0x baseline for natural flow.
- Increase speed for quick notifications.
- Decrease speed for complex instructions.
Evaluating Clarity in Noise
Play generated audio alongside background noise typical of your app’s usage context. Assess whether the synthetic speech remains intelligible. This ensures the voice performs reliably in real-world conditions without requiring additional audio processing.
- Test against common ambient sounds.
- Check volume consistency across speeds.
- Verify no audio artifacts at limits.

Common Mistakes
Developers often assume the API detects language automatically, leading to awkward phrasing in non-English Latin texts. Setting the language parameter explicitly ensures correct pronunciation and intonation. Without it, the system defaults to English, which may misinterpret words from other languages.
Another frequent error is selecting voices solely by gender. This ignores the specific tonal qualities, such as Luna’s storytelling softness or Felix’s energetic promo style, which better match your app’s content genre. Gender is a secondary filter, not a primary selector.
Users also overlook speed limits. The API supports speeds from 0.25 to 4.0, but extreme speeds can reduce clarity. Testing your specific text at various speeds helps find the optimal balance between generation speed and listener comprehension.
Ignoring Language Settings
The API guesses the language from the script if not specified. For Latin-script text that is not English, this guess is often incorrect. Always provide the language parameter with the correct ISO code to ensure accurate phonetic rendering.
This mistake is common when localizing apps quickly. Developers send French or German text without the language tag, resulting in an English accent that undermines the user experience. Explicit configuration is required for consistent quality.
- Set
languagefor all non-English Latin text. - Default detection may fail for mixed-language inputs.
- Incorrect accents reduce perceived realism.
Choosing Based on Gender Only
Gender is a basic categorization, but voice character matters more. Aria offers a clear narrator tone, while Orion provides a deep documentary style. Selecting based on gender alone might assign a bright voice to a serious news app, creating tonal dissonance.
Consider the emotional context of your content. A friendly conversational tone like Milo’s suits chatbots, whereas a neutral news anchor style like Atlas’s fits informational updates. Matching tone to genre improves user engagement significantly.
- Match tone to content genre, not just gender.
- Aria suits narration; Orion suits documentaries.
- Milo works well for conversational interfaces.
Overlooking Speed Limits
The speed parameter ranges from 0.25 to 4.0. Setting it too high can make speech unintelligible, while too low may feel sluggish. Test your text at various speeds to find the optimal listening experience.
Speed adjustments also affect generation time. Higher speeds generate audio faster but may compromise clarity. Ensure your users can adjust speed settings within your app for accessibility and preference.
- Range: 0.25 to 4.0.
- High speeds may reduce clarity.
- Test optimal speed for your content.
Integration Tips
Efficient API consumption requires strategic data management. Since input text is not stored but audio is returned in the response, caching generated files locally prevents redundant charges for identical inputs. This is particularly effective for static app content like menus or help sections where the voice output remains constant.
Implementing local storage reduces API calls, preserving your character limit for dynamic user-generated content. Ensure your cache keys account for voice, speed, and format parameters to avoid serving mismatched audio files.
Caching Audio Files
Store generated audio files locally using unique keys based on input text, voice, and parameters. This avoids re-synthesizing identical strings, saving characters and processing time.
Invalidation strategies should consider parameter changes. If speed or format changes, the cached file is no longer valid and requires a new request.
- Use MD5 hashes of parameters as cache keys.
- Clear cache on app updates if voice models change.
- Prioritize caching for static UI elements.
Batch Processing Text
Split long texts into segments of up to 4096 characters to fit within a single request. Join the resulting audio files client-side or via a simple concatenation script.
This approach maintains the natural flow of the audio while adhering to the API's input constraints.
Process segments sequentially or in parallel depending on your rate limit of 60 requests per minute.
- Max 4096 characters per request.
- Concatenate MP3 or WAV outputs.
- Monitor queue times under load.
Error Handling for Voices
Handle API errors gracefully to avoid user-facing glitches. If a voice parameter is invalid or the rate limit is exceeded, retry with exponential backoff.
Failed requests are free, so retrying does not incur extra character costs if the error is consistent.
Log errors to identify if a specific voice or format causes consistent failures.
- Implement retry logic for 429 Too Many Requests.
- Validate voice IDs before sending.
- Log failed synthesis attempts for debugging.
Cost Implications
Billing depends strictly on input characters, including spaces and punctuation. At $0.01 per 1,000 characters, precise text length management directly controls costs. Failed requests incur no charge, providing a safety net during development.
Character Count Variations
Each request supports up to 4,096 characters. Longer texts require splitting and joining audio files, which may increase processing time but does not alter the per-character rate. Ensure your input text stays within this limit to avoid unnecessary segmentation overhead.
Format Size Differences
Choose your response_format wisely based on your deployment needs. MP3 offers smaller file sizes suitable for web delivery, while WAV or FLAC provides higher quality for local playback. PCM is uncompressed and largest, whereas OPUS offers efficient compression. The character count remains the billing metric, regardless of the output format size.
Trial Usage Strategy
New accounts signing in with Google receive $0.50 in trial credit, valid for 7 days. This allows for approximately 50,000 characters of testing. Accounts created with email and password start without this trial credit. Use this period to validate voice selection and integration logic before committing to USDT top-ups.
Future Voice Updates
Lumivoice maintains a curated catalog of eight licensed voices derived from the LibriTTS-R dataset. Future updates may introduce additional presets from this source material to expand stylistic variety without altering the core engine.
We monitor usage patterns and developer feedback to prioritize new voice styles that align with indie app needs, such as specialized narrative or technical tones. Changes are implemented only when they add distinct value to the existing set.
Potential New Voices
New voices will be added based on clear demand and availability within the licensed LibriTTS-R archive. We do not promise specific timelines or new voice releases, as the current eight presets cover the majority of indie use cases effectively.
Community Feedback Loop
Developers can submit feature requests via our support email. We review these requests quarterly to determine if a new voice style warrants development. Feedback directly influences which archival recordings are prioritized for voice preset conversion.
- Email support for feature requests
- Quarterly review cycle
- Priority based on archival availability
No Custom Cloning Promise
We do not offer custom voice cloning or uploading of proprietary voice samples. All voices remain based on the static licensed recordings. This ensures consistent quality and predictable licensing terms for all users without the complexity of custom model training.
Summary of Choices
Lumivoice offers eight distinct voices across two genders, each optimized for specific content types. This guide consolidates those options to help indie developers make immediate, informed decisions without unnecessary evaluation.
Quick Reference Table
The table below maps each voice to its primary use case, ensuring rapid selection for your application's audio needs.
- Aria: Fiction narration
- Nova: Promotional content
- Luna: Storytelling
- Sage: Educational instructions
- Felix: Energetic promos
- Milo: Conversational interfaces
- Atlas: News reporting
- Orion: Documentaries
Best Fit Recommendations
For narrative-driven apps, select Aria or Luna to maintain listener engagement. Use Sage or Milo for instructional guides and conversational agents where clarity is paramount. Felix suits short promotional clips, while Atlas and Orion handle structured reporting or documentary styles effectively.
Final Selection Checklist
Before finalizing your voice choice, verify the target language is supported in our 23-language list. Ensure the selected voice aligns with your content genre and that speed adjustments between 0.25x and 4.0x will not distort clarity. Confirm that the audio format (mp3, wav, etc.) matches your player’s capabilities.
- Check language support
- Verify genre fit
- Set speed limits
- Confirm format compatibility
Listen to the 8 voices
Each sample below was generated by this API from the same sentence.
23 languages
Any voice speaks any of these languages.
Sources and further reading
- Wikipedia
- TRON developer hub: TRC-20
- curl documentation
- Creative Commons Attribution 4.0
- MDN: Web Speech API
- FFmpeg documentation
- LibriTTS-R corpus (OpenSLR 141)
- W3C WAI: making audio and video accessible
- Tether (USDT)
- ISO 639 language codes
- LibriTTS-R paper (arXiv 2305.18802)
- OpenAI API reference: create speech
