Frequently asked questions
What is the difference between Aria and Luna?
Aria is a warm, clear narrator voice, while Luna is soft and designed for storytelling. Both are female presets available in the API. You can select either via the voice parameter in your POST request to https://api.getpredictz.com/v1/audio/speech based on whether your content requires a direct narrative or a gentle, story-like tone.
Which voice is best for a news app?
Atlas is the recommended voice for news applications. Described as neutral, it suits news delivery well. You can also consider Sage for instructional content or Felix for energetic promos. Select the voice via the voice field in your API request. All voices support the languages your app requires, ensuring consistency across your voice interface.
Can a female voice speak Spanish?
Yes. All eight voice presets, including female voices like Aria, Nova, Luna, and Sage, can speak Spanish. You can specify Spanish using the optional language parameter with the ISO code. If omitted, the system guesses the language, but setting it explicitly ensures accurate pronunciation for your chatbot or voice assistant.
How do I make the voice sound faster?
Adjust the speed parameter in your API request. The supported range is 0.25 to 4.0. Setting a value above 1.0 increases the playback speed. This applies to all voices and formats. Ensure your text input does not exceed 4096 characters, as requests are processed sequentially and may queue under load.
Does the API support emotion tags?
No, the API does not support emotion tags. It also lacks SSML, phoneme dictionaries, and custom pronunciation lexicons. You control tone by selecting one of the eight specific voice presets (e.g., Felix for energetic, Luna for soft) and adjusting the speed. Each generated file includes an inaudible watermark marking it as synthetic speech.
Are the voices robotic or natural?
The voices are based on licensed recordings from LibriTTS-R and LibriVox public-domain audiobook readings. They are designed to sound natural for chatbots and voice assistants. Each audio file contains an inaudible watermark indicating it is synthetic speech. The voices are not imitations of real celebrities or politicians, ensuring consistent, natural-sounding output.
Can I use the same voice for all languages?
Yes, any of the eight voice presets can speak any of the 23 supported languages. You do not need different voices for different languages. Simply set the language parameter to the desired ISO code. If unspecified, the system guesses the language, defaulting to English for Latin-script text, so explicit setting is recommended for non-English languages.
How do I handle numbers in text?
The API does not support SSML or custom pronunciation lexicons, so you cannot explicitly control number pronunciation. You must format numbers as text in your input string. The system guesses the language based on the script. For non-English Latin text, set the language parameter to ensure correct interpretation of numbers and punctuation.
Is there a limit on voice changes?
There is no limit on how often you can change voices between requests. You can switch voices dynamically for each call to https://api.getpredictz.com/v1/audio/speech. The only limits are the 4096 character input per request and the rate limit of 60 requests per minute per API key. Each request is processed independently.
Do voices sound different in MP3 vs WAV?
The voice synthesis engine is the same for all formats, including mp3, wav, opus, flac, and pcm. The difference lies in compression and file size, not vocal quality. MP3 is compressed, while WAV is uncompressed. Choose based on your storage and bandwidth needs. The inaudible watermark is present in all generated audio files.
Can I clone a celebrity voice?
No, the API does not offer custom voice cloning. You cannot upload your own voice or clone celebrities. The catalog consists of eight licensed presets based on LibriVox recordings. If you need a specific voice, you must choose from the available options: Aria, Nova, Luna, Sage, Felix, Milo, Atlas, or Orion.
What is the watermark in the audio?
Every generated audio file carries an inaudible watermark that marks it as synthetic speech. This applies to all voices and formats. The watermark is embedded during synthesis and is not visible or audible to users. It serves to identify the audio as AI-generated, which is standard for synthetic speech outputs from this API.
How do I test voices before paying?
Sign in with Google to receive a $0.50 free trial credit, valid for 7 days. This covers about 50 000 characters. Use your API key to make requests to the speech endpoint. If you use email and password, you start without trial credit. Failed requests are free, so you can test error handling without cost.
Is SSML supported for emphasis?
No, the API does not support SSML. You cannot use tags for emphasis, pauses, or pronunciation. Control the output by selecting the appropriate voice preset and adjusting the speed parameter. The system guesses language from the script, so set the language field explicitly for non-English text to ensure accurate synthesis without SSML instructions.
Understanding voice personas
Choosing the right voice persona is critical for chatbots and voice assistants. Predictz Voice offers eight distinct presets, each designed for specific interaction styles. Understanding these personas helps you align your assistant’s personality with user expectations.
Warm and clear narrators
Aria provides a warm, clear narration style ideal for storytelling or guided tours. Luna offers a softer, storytelling tone that feels intimate and engaging. These voices work well for onboarding sequences or calm user experiences.
- Aria: Warm, clear narrator
- Luna: Soft, storytelling tone
Neutral and professional tones
Atlas delivers a neutral, news-like delivery suitable for factual updates or weather reports. Sage provides a calm, instructional voice perfect for technical support or step-by-step guides. These voices ensure clarity without emotional bias.
- Atlas: Neutral, news style
- Sage: Calm, instructional
Energetic and upbeat styles
Nova brings a bright, upbeat energy to promotions or quick interactions. Felix offers an energetic, promo-style voice that captures attention quickly. These presets are effective for alerts, reminders, or promotional content.
- Nova: Bright, upbeat
- Felix: Energetic, promo style
Matching voice to brand identity
Matching voice to brand identity ensures your voice assistant feels like a natural extension of your product. The right persona builds trust and aligns with user expectations, whether you are guiding customers through a transaction or providing informational support.
Friendly conversational styles
For chatbots and customer support interfaces, voices like Milo or Felix create an approachable atmosphere. Milo offers a friendly, conversational tone ideal for everyday interactions, while Felix brings energetic enthusiasm suitable for promotional content or quick updates. These presets help reduce user friction by sounding human without being overly formal.
- Milo: Friendly, conversational tone for general assistance
- Felix: Energetic style for promos and announcements
Deep documentary voices
When delivering news, facts, or serious information, Atlas provides a neutral, professional delivery that commands attention. Orion’s deep, documentary-style voice adds gravitas to longer explanations or educational content. These voices ensure clarity and authority, making complex data easier to digest.
- Atlas: Neutral, professional tone for news and facts
- Orion: Deep, documentary style for educational content
Soft storytelling tones
For narrative-driven apps, such as audiobooks or guided meditations, Luna’s soft, storytelling voice creates an immersive experience. Aria offers a warm, clear narration style that works well for instructional content requiring a gentle touch. These voices prioritize emotional resonance and consistency, ensuring users remain engaged without feeling rushed by synthetic delivery.
- Luna: Soft, storytelling tone for narratives
- Aria: Warm, clear narration for instructional content
Technical considerations for chatbots
Latency is critical for conversational flow. Short phrases take a few seconds to generate, while longer texts process at roughly real-time speed (25 characters per second). Requests are queued sequentially, so under load, generation may delay. This sequential processing means consistent timing depends on current server volume.
Clarity ensures the voice assistant remains understandable even in background noise. The engine’s design prioritizes natural speech patterns over extreme articulation. For noisy environments, selecting a voice like Felix or Nova, which have energetic or bright profiles, may improve intelligibility.
Latency and generation speed
Latency is critical for conversational flow. Short phrases take a few seconds to generate, while longer texts process at roughly real-time speed (25 characters per second). Requests are queued sequentially, so under load, generation may delay. This sequential processing means consistent timing depends on current server volume.
Clarity ensures the voice assistant remains understandable even in background noise. The engine’s design prioritizes natural speech patterns over extreme articulation. For noisy environments, selecting a voice like Felix or Nova, which have energetic or bright profiles, often improves intelligibility without requiring additional audio post-processing.
- Short phrases: few seconds generation time
- Long texts: ~25 characters per second
- Sequential processing; queueing under load
Clarity in noisy environments
Clarity ensures the voice assistant remains understandable even in background noise. The engine’s design prioritizes natural speech patterns over extreme articulation. For noisy environments, selecting a voice like Felix or Nova, which have energetic or bright profiles, often improves intelligibility without requiring additional audio post-processing.
Consistency across languages
Consistency across languages is maintained by using the same voice presets for all 23 supported languages. Whether your user speaks English, Spanish, or Japanese, the timbre and style remain identical. This uniformity helps maintain brand identity regardless of the user's locale.
Auto-detection works for most scripts, but setting the language parameter explicitly ensures accurate pronunciation for non-English Latin text. This prevents the engine from defaulting to English phonemes for words like 'café' or 'naïve', ensuring a consistent listening experience.
- Same voice presets across 23 languages
- Explicit
languagecode for non-English Latin text - Uniform timbre maintains brand identity
Testing voice performance
Testing voice performance is essential before committing to a voice assistant design. With $0.50 of free trial credit, you can generate approximately 50,000 characters without financial risk. This allows you to validate that the audio output meets your quality standards before scaling up with paid top-ups via USDT (TRC-20).
Use this credit to generate diverse sample phrases that reflect your typical user interactions. Evaluate the naturalness of the synthesis. This iterative process helps you select the best voice preset for your specific use case.
Use the free trial credit
New accounts signing in with Google receive $0.50 of credit, valid for 7 days. This covers about 50,000 characters, allowing you to test multiple voices and languages without immediate payment. If you use email and password, you start without trial credit and must top up first.
- $0.50 credit for Google sign-ups
- Valid for 7 days
- Covers ~50,000 characters
Generate sample phrases
Create test inputs covering your most common dialogue patterns. Include questions, commands, and conversational fillers. Use the API to generate audio for each phrase, checking that the selected voice handles punctuation and pauses correctly. Ensure the output matches the expected tone for your brand identity.
- Test questions, commands, and fillers
- Check punctuation handling
- Verify tone consistency
Evaluate naturalness
Listen to the output. Compare different voices like Aria for clarity or Felix for energy. Use the speed parameter (0.25–4.0) to adjust pacing if the default feels too fast or slow. Ensure the audio is clear and understandable in your target environment.
- Listen for robotic artifacts
- Adjust speed with the speed parameter
- Compare voices for brand fit
Language support and accents
Predictz Voice supports 23 languages, covering major global markets including English, Spanish, German, Japanese, and Arabic. This range ensures your voice assistant can communicate effectively with a diverse user base without needing separate voice engines for different regions.
23 supported languages
The supported list includes 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, allowing you to maintain a consistent brand identity across different linguistic markets.
Any voice speaks any language
Unlike many TTS providers that tie specific voices to specific languages, Predictz allows any of the 8 presets to speak all 23 languages. You can use the warm Aria voice for English, French, and Spanish, or the deep Orion voice for Russian and German, ensuring tonal consistency regardless of the language spoken.
Auto-detection vs explicit code
If the language field is omitted, the API guesses the language from the text script. Latin script text defaults to English. For non-English Latin-script text, you must explicitly set the language parameter to ensure correct pronunciation. If a language is not supported, the API will return an error rather than attempting a guess.
Speed adjustments for UX
Adjusting speech speed directly impacts user experience in voice interfaces. The API accepts a speed parameter ranging from 0.25x to 4.0x, allowing you to fine-tune delivery for different contexts. Slower rates improve clarity for complex instructions, while faster rates suit quick confirmations.
Optimizing for speed involves balancing naturalness with efficiency. Since generation takes roughly one second per second of audio, faster playback reduces perceived wait times. However, extreme speeds may sacrifice intelligibility, so testing with your specific content is essential.
Range from 0.25x to 4.0x
The API supports a wide range of speeds, enabling precise control over voice pacing. This flexibility allows your voice assistant to adapt its tone dynamically, whether delivering calm narration or rapid-fire updates.
You can set the speed per request, ensuring each response matches the intended urgency. This granular control helps maintain a consistent brand voice across varying interaction types.
- Minimum speed: 0.25x for slow, deliberate speech
- Maximum speed: 4.0x for rapid delivery
- Default speed: 1.0x for natural pacing
Optimizing for quick responses
For chatbots, faster speeds can make interactions feel more responsive. Use higher speeds for short confirmations or status updates where detailed enunciation is unnecessary.
However, consider that generation time is tied to audio length. Faster playback does not reduce generation time, but it can make the interaction feel quicker to the user.
- Use 1.2x–1.5x for quick confirmations
- Reserve 2.0x+ for very short utterances
- Avoid speeds above 3.0x for complex data
Balancing clarity and speed
Clarity is paramount for usability. While faster speeds save time, they can reduce understanding, especially for non-native speakers or in noisy environments. Test your specific use case to find the optimal balance.
Start with default speeds and adjust based on user feedback. Monitor drop-off rates or confusion indicators to determine if your voice assistant is speaking too fast or too slow.
- Test with diverse user groups
- Monitor comprehension metrics
- Adjust based on context
Audio format impact on quality
The response_format parameter determines the binary structure of your audio payload. Choosing MP3 provides a balanced trade-off between file size and compatibility, while WAV delivers uncompressed lossless audio for maximum fidelity.
Select OPUS to minimize bandwidth consumption, ideal for low-latency chatbot interactions. Use PCM if your downstream processing pipeline requires raw, uncompressed PCM data.
MP3 vs WAV differences
MP3 files are significantly smaller, reducing data transfer costs and improving load times for voice assistants. WAV files retain every audio sample, offering the highest potential quality but requiring more storage and bandwidth for the same duration.
- MP3: Compressed, smaller file size, broad device support.
- WAV: Uncompressed, larger file size, lossless quality.
OPUS for bandwidth efficiency
OPUS is optimized for efficient transmission over networks. It maintains good audio quality at lower bitrates, making it suitable for mobile apps or environments with limited connectivity.
- Ideal for mobile voice assistants.
- Reduces data usage for frequent short queries.
PCM for raw processing
PCM provides uncompressed audio samples directly. This format is best when you need to apply real-time audio effects or feed the audio into another processing engine without decoding overhead.
- No compression artifacts.
- Immediate availability for audio analysis.

Licensing and usage rights
All Predictz Voice audio is generated from licensed recordings sourced from LibriTTS-R, which contains public-domain audiobook readings from LibriVox under the CC BY 4.0 license. This foundation ensures that your voice assistant’s output is built on legally clear, high-quality source material without hidden royalty burdens.
You can use the generated audio commercially without additional attribution requirements beyond the license terms. The API returns the audio file directly, allowing seamless integration into your chatbot or voice interface. There are no restrictions on volume or specific use cases, making it suitable for both startup MVPs and scaled products.
LibriVox public domain source
The voice models are derived from LibriTTS-R, a dataset built from LibriVox public-domain audiobook readings. This ensures that the underlying vocal characteristics are free from restrictive copyright claims, providing a clean legal baseline for your voice products.
Since the source material is public domain, you avoid the complexity of negotiating individual voice actor rights. The data is pre-licensed under CC BY 4.0, simplifying compliance for startups focused on speed and reliability.
- Sourced from LibriTTS-R dataset
- Based on LibriVox public-domain readings
- Licensed under CC BY 4.0
Commercial usage allowed
You may use the synthesized audio for commercial purposes, including in paid applications, customer service bots, and marketing materials. There are no per-unit fees or revenue caps tied to the license itself.
The generated files are yours to distribute. Whether you are building a free utility app or a subscription-based voice assistant, the voice assets remain usable without ongoing royalty payments or attribution clauses beyond standard requirements.
Watermark in synthetic audio
Every generated audio file contains an inaudible watermark that identifies the speech as synthetic. This does not affect playback quality or user experience but provides a clear marker of origin for your records.
The watermark is embedded in the audio data itself, ensuring that the synthetic nature of the voice is preserved even if the file is distributed or archived. It serves as a transparent indicator of AI-generated content.
Handling edge cases in speech
Edge cases often disrupt voice assistant flow. Incorrect pronunciation of numbers or abrupt pauses can make synthetic speech sound robotic. Understanding how the engine handles these details ensures a smoother user experience.
Numbers and abbreviations
The engine reads text literally. "100" becomes "one hundred," not "hundred." Use "100 percent" if you want that specific phrasing. Abbreviations like "Dr." or "Mr." are read as written unless spelled out. Test common industry jargon to ensure clarity.
- Spell out complex acronyms for clarity
- Test date formats (e.g., "12/25")
- Verify currency symbols read correctly
Punctuation and pauses
Punctuation controls rhythm. Commas create short pauses; periods create longer stops. No SSML tags are available, so use punctuation strategically. Exclamation marks add slight emphasis. Ellipses (...) create a trailing effect.
- Use commas for natural breaks
- Use periods for full stops
- Avoid excessive punctuation for clarity
Character limit handling
Each request supports up to 4096 characters. Longer texts must be split into parts. The API returns separate audio files for each part. You must join these files on the client side or via your backend logic to create a continuous audio file.
- Split text into 4096-char chunks
- Concatenate audio files sequentially
- Monitor total character count per minute
Monitoring usage and costs
Predictive cost management relies on tracking character consumption against your balance. Since billing occurs per million characters, understanding your usage helps avoid unexpected service interruptions during high-traffic periods.
Monitoring rate limits ensures your voice assistant remains responsive. With a base limit of 10 requests per minute, scaling requires careful queue management or top-ups to maintain smooth user experiences.
Track characters consumed
Each request consumes characters based on input length, including spaces and punctuation. Your dashboard displays remaining balance and trial credits, allowing precise budgeting for long-form narration or frequent short queries.
Failed requests incur no charge, so you can safely test edge cases without impacting your primary usage metrics. Regular monitoring prevents overspending during peak deployment phases.
- Characters include spaces and punctuation
- Failed requests are free
- Balance visible in dashboard
Monitor rate limits
Standard accounts allow 60 requests per minute, rising to 10 before the first top-up. Exceeding these limits queues subsequent requests, potentially increasing latency for time-sensitive voice responses.
A single active API key governs all requests. Issuing a new key instantly invalidates the old one, so ensure your application switches keys seamlessly if rotating for load distribution.
Optimize text length
Split texts exceeding 4,096 characters into smaller segments to ensure consistent synthesis. Joining audio files from multiple requests maintains continuity while respecting API constraints.
Efficient text formatting reduces character count, lowering costs. Removing unnecessary whitespace or simplifying phrasing can significantly extend your usable character budget without compromising clarity.
Integration with voice frameworks
Predictz Voice integrates into existing voice ecosystems by returning standard audio files, simplifying playback logic for chatbots and assistants.
This approach ensures consistent playback across devices without managing WebSocket connections or partial audio buffering.
Web Speech API compatibility
Since the API returns standard audio files, you can easily feed the response into the Web Speech API or standard HTML5 audio elements. This allows for straightforward integration into web-based voice assistants without requiring custom media handlers.
- Supports MP3, WAV, OPUS, FLAC, and PCM formats for broad browser compatibility.
- No custom streaming protocol required; standard HTTP POST returns complete audio files.
- Works with any JavaScript audio library that accepts Blob or ArrayBuffer inputs.
Mobile app integration
Mobile applications can cache generated audio files locally to reduce latency for frequently used phrases. The consistent request shape allows you to reuse existing HTTP clients for both text and audio requests.
- Ideal for on-demand voice responses in iOS and Android apps.
- Low bandwidth usage with OPUS format for efficient mobile data transfer.
- Simple integration with native audio players using standard file URLs.
Server-side generation
Generate audio on your server before sending it to the client, ensuring a consistent user experience. This method hides API key exposure and allows for pre-processing of text before synthesis.
- Secure API key management with server-side authentication.
- Pre-generate audio for static content to improve response times.
- Simplify client-side logic by delivering ready-to-play audio files.
Future proofing your voice stack
Predictz Voice relies on a stable API structure compatible with standard OpenAI SDKs, ensuring your integration remains resilient as your startup scales. The underlying engine uses two models, "tts-1" and "tts-1-hd," allowing you to deploy voice interfaces with confidence.
While the core endpoint remains steady, the service prioritizes reliability over frequent feature overhauls. You can integrate today knowing the request shape and authentication methods will remain consistent, reducing the technical debt associated with constant API migrations in fast-paced development cycles.
API stability guarantees
The API maintains a consistent request shape aligned with industry standards, ensuring that your existing codebase remains functional without unexpected breaking changes. This stability allows startups to focus on building user experiences rather than rewriting integration logic.
Authentication via bearer tokens and standard JSON bodies ensures predictable behavior across different environments, from development sandboxes to production servers.
- Consistent JSON request structure
- Standard bearer token authentication
- No breaking schema changes
Model updates
The service operates on two primary models: "tts-1" and "tts-1-hd." These models share the same engine and pricing structure, providing a straightforward upgrade path without complex migration strategies. You can switch between standard and high-quality audio by simply changing the model parameter.
Future improvements to the synthesis engine will be rolled out transparently. Since the voice presets remain constant, any underlying quality enhancements will automatically benefit your existing integrations without requiring code changes.
- Two stable model options
- Transparent engine improvements
- No mandatory migration periods
Voice catalog expansion
The current catalog offers eight distinct voice presets, ranging from warm narrators to deep documentary tones. This fixed set ensures consistent brand identity across your application, as voices do not change unexpectedly between requests.
As the platform evolves, new voices may be added to the /v1/voices endpoint. Existing integrations will continue to function with the original eight voices, while new options become available for fresh deployments or A/B testing without disrupting live services.
- Eight consistent voice presets
- New voices added via API
- Existing voices remain stable
Listen to the 8 voices
Each sample below was generated by this API from the same sentence. This allows a single voice identity to maintain brand consistency across global markets without needing to retrain or swap voice assets for each new region.
23 languages
Any voice speaks any of these languages.
