Cloud Text-to-Speech API for Reading Text Aloud

By AudioLex Access editorial team · Updated 2026-11-19

This guide details the technical specifications and performance characteristics of the AudioLex Access text-to-speech API, designed specifically for developers integrating audio playback into accessibility and dyslexia-focused reading applications. You will find concrete data on voice profiles, language support, and format options to help you select the optimal configuration for your users.

We analyze eight distinct voice presets across male and female categories, highlighting their suitability for narrative, instructional, and news content. The page covers practical considerations including latency, CPU usage, and file size impacts, ensuring you can build a reliable, compliant reading experience without managing the underlying synthesis infrastructure.

Cloud Text-to-Speech API for Reading Text Aloud

Frequently asked questions

Which voice is the clearest?

Aria is described as a warm, clear narrator, making it suitable for clear reading. Luna is also noted as soft and storytelling-oriented. Clarity depends on the text content and speed settings. You can test any of the eight available voices via the API to determine which best suits your specific accessibility use case.

Can I choose a voice for each chapter?

Yes. Each API request allows you to specify a different voice from the eight presets. You can split your text into chapters, send separate requests for each, and receive individual audio files. This allows you to assign distinct voices to different sections of your content programmatically.

Do all voices speak all languages?

Yes. Any of the eight voices can speak any of the 23 supported languages. If the language is not specified, it is guessed from the script. For non-English Latin text, you should explicitly set the language parameter to ensure accurate pronunciation across all voice presets.

Is there a difference in cost between voices?

No. The model parameter accepts either tts-1 or tts-1-hd, which use the same engine and price. All eight voices are generated using these models. The cost is $10 per 1 million characters regardless of which voice you select for the synthesis request.

Can I clone a custom voice?

No. The service does not offer custom voice cloning or the ability to upload your own voice. You must choose from the eight available presets: Aria, Nova, Luna, Sage, Felix, Milo, Atlas, or Orion. These are based on licensed recordings from LibriTTS-R and LibriVox.

Do voices have emotion tags?

No. The API does not support emotion tags or custom pronunciation lexicons. Voices are described by general tones like warm, bright, or calm. You control the delivery primarily through the speed parameter, which ranges from 0.25 to 4.0, allowing for faster or slower reading rates.

Which voice is best for children?

Luna is soft and storytelling-oriented, while Nova is bright and upbeat. These voices are suitable for children. Felix is energetic and promo-focused. Since there is no specific child preset, test these voices with your target text to determine the most appropriate tone for your audience.

Can I change voices during playback?

The API returns a single audio file per request. To change voices during playback, you must generate separate audio files for each voice segment and manage the switching in your application player. The API does not support dynamic voice changes within a single continuous audio stream.

Are the voices human or synthetic?

The voices are synthetic. Every generated file carries an inaudible watermark marking it as synthetic speech. The voices are based on licensed recordings from LibriTTS-R and LibriVox, but the output is generated by the TTS engine, not a direct playback of the original human recordings.

Do voices sound different in WAV vs MP3?

The voice tone remains the same across formats. The difference is in the file format: MP3, WAV, OPUS, FLAC, or PCM. WAV and FLAC are lossless, while MP3 and OPUS are compressed. Choose based on your storage and bandwidth requirements, not for voice quality variation.

Can I use multiple voices in one app?

Yes. Your application can send multiple requests, each specifying a different voice from the eight presets. You can mix voices within a single project by generating separate audio files for each segment. This allows for dynamic voice assignment based on content type or user preference.

Is there a limit on voice changes?

There is no specific limit on how often you can change voices between requests. You are limited by the rate limit of 60 requests per minute per key. Each request can specify a different voice, allowing frequent changes as long as you stay within the quota.

Do voices work offline?

No. This is a cloud API. You send text via HTTP POST to the endpoint and receive an audio file. The synthesis happens on the server. You need an internet connection to make requests and receive the generated audio files for playback in your application.

How do I test voice quality?

Send a test request to the API with sample text and your chosen voice. The response will contain the audio file. You can listen to it to assess clarity and tone. Use the free trial credit of $0.50 to test multiple voices and formats before committing to paid usage.

Female Voices for Narrative Reading

Sound wave illustration for: Female Voices for Narrative Reading

AudioLex Access provides three female voice models optimized for narrative reading workflows. These presets deliver distinct tonal qualities suitable for accessibility applications, ensuring clear audio output for diverse user needs.

Each voice maintains consistent performance across supported languages, allowing developers to select the most appropriate tone for their specific content

Aria: Warm and clear

Aria delivers a warm, clear narration style ideal for general reading applications. Its balanced tone reduces listener fatigue during extended sessions, making it suitable for standard text-to-speech workflows.

Users can adjust the speed between 0.25 and 4.0 to match individual reading preferences without compromising audio clarity.

  • Ideal for general accessibility reading
  • Warm, clear narrative tone
  • Consistent clarity at variable speeds

Luna: Soft storytelling

Luna offers a soft, storytelling voice designed for narrative content. The gentle tone supports immersive reading experiences, particularly for users who prefer a less directive audio presentation.

Like all voices, Luna supports 23 languages and maintains consistent performance regardless of the selected output format.

  • Soft, engaging storytelling tone
  • Supports immersive reading styles
  • Available in 23 languages

Nova: Bright and upbeat

Nova provides a bright, upbeat delivery suitable for energetic content or younger audiences. Its distinct tonal profile ensures high intelligibility while maintaining a positive auditory character.

Developers can utilize Nova for educational materials or promotional texts where a lively tone enhances user engagement without requiring custom voice design.

  • Bright, upbeat delivery style
  • Suitable for educational content
  • High intelligibility for diverse audiences

Male Voices for Instructional Content

Sound wave illustration for: Male Voices for Instructional Content

Male voices provide stable, authoritative audio for instructional and narrative reading applications. Sage, Milo, and Atlas offer distinct tonal profiles suitable for diverse accessibility needs.

These presets handle up to 4096 characters per request. Longer texts must be split into parts and joined by the developer.

Sage: Calm and instructional

Sage delivers a calm, instructional tone ideal for step-by-step guides. Its steady pace aids comprehension for users with processing delays.

  • Optimized for clear instruction delivery
  • Steady rhythm reduces cognitive load

Milo: Friendly and conversational

Milo offers a friendly, conversational style suitable for informal reading. It maintains engagement while preserving clarity for dyslexic readers.

  • Conversational tone for engagement
  • Balanced energy for long texts

Atlas: Neutral and news-like

Atlas provides a neutral, news-like delivery for factual content. Its consistent pitch ensures accurate information transfer without emotional bias.

  • Neutral tone for factual content
  • Consistent pitch for accuracy

Voice Characteristics Comparison

Sound wave illustration for: Voice Characteristics Comparison

AudioLex Access offers eight distinct voice presets, each engineered for specific auditory profiles. Female voices like Aria provide a warm, clear narrator style, while Luna offers a soft, storytelling tone. Male options range from Sage’s calm instructional delivery to Atlas’s neutral, news-like clarity.

These variations allow developers to match the auditory experience to the content’s nature. For instance, Milo’s friendly conversational style suits interactive learning modules, whereas Orion’s deep documentary tone works well for factual reporting. The choice depends on the desired listener engagement level.

Pitch and tone differences

Each voice exhibits unique pitch and tonal qualities. Aria’s warm clarity contrasts with Nova’s bright, upbeat energy. Sage delivers calm instruction, while Felix projects energetic promotional tones. Atlas remains neutral, and Orion provides a deep, documentary-style depth.

These differences influence listener retention and comprehension. Higher-pitched voices like Nova may feel more engaging for younger audiences, while lower tones like Orion convey authority. Developers should select voices that align with the content’s emotional weight without relying on dynamic shifts.

  • Aria: Warm, clear narrator
  • Nova: Bright, upbeat
  • Luna: Soft, storytelling
  • Sage: Calm, instructional
  • Felix: Energetic, promo
  • Milo: Friendly, conversational
  • Atlas: Neutral, news
  • Orion: Deep, documentary

Clarity for dyslexic readers

Clarity is paramount for accessibility. Voices like Aria and Sage offer clear articulation, reducing cognitive load for dyslexic readers. Consistent pacing and distinct phoneme enunciation aid comprehension. The API’s models prioritize clear diction over stylistic flair.

Avoiding overly expressive voices prevents distraction. Neutral voices like Atlas provide a stable baseline, ensuring that the text’s meaning remains the focus. Developers can test voices with sample texts to verify clarity for their specific user base.

  • Clear articulation reduces cognitive load
  • Consistent pacing aids comprehension
  • Neutral tones minimize distraction
  • Test voices with sample texts

No emotion tags

The API does not support SSML or emotion tags. Voices deliver a consistent tone without dynamic emotional shifts. This simplifies integration but requires manual selection of the appropriate preset to match the content’s mood. For example, use Luna for soft stories and Felix for energetic promos.

This limitation ensures predictable audio output. Developers cannot dynamically adjust emotion via API parameters. Instead, they must choose the most suitable voice for the entire text or switch voices at paragraph boundaries. This approach maintains consistency and reduces processing complexity.

  • No SSML support
  • No dynamic emotion tags
  • Consistent tone per voice
  • Manual voice selection required

Performance in Different Languages

Sound wave illustration for: Performance in Different Languages

AudioLex Access supports 23 languages, including English, Spanish, French, German, and Japanese. Any voice can speak any of these languages, allowing developers to build consistent reading experiences across diverse user bases without managing multiple voice models. The API automatically detects the language from the input script, though explicit language codes are recommended for non-English Latin text to ensure accuracy.

English proficiency

English is the default language for Latin-script text when no language field is specified. The voices deliver clear, standard pronunciation suitable for accessibility applications, ensuring high comprehension for readers with dyslexia or visual impairments. Input text is processed character-by-character, with spaces and punctuation counting toward the 4,096-character limit per request.

  • Default detection for Latin scripts
  • Clear standard pronunciation
  • Accurate character counting for billing

Multilingual voice versatility

The eight preset voices handle all supported languages without switching engines. This uniformity simplifies integration, as developers use the same voice parameter regardless of the target language. The system supports Arabic, Chinese, Danish, Dutch, Finnish, Greek, Hebrew, Hindi, Italian, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Swahili, Swedish, and Turkish.

  • Single voice model for 23 languages
  • No engine switching required
  • Consistent audio quality across languages

Accent consistency

Voices maintain consistent accent characteristics regardless of the language spoken. For example, Aria remains a warm, clear narrator whether reading English or French. This predictability helps users build familiarity with the voice’s tone and pacing, which is critical for long-form reading applications where listener fatigue is a concern.

  • Consistent vocal timbre across languages
  • Predictable tone for user familiarity
  • No accent drift during synthesis

Technical Requirements per Voice

Sound wave illustration for: Technical Requirements per Voice

The TTS engine processes requests sequentially, meaning CPU load correlates directly with queue depth rather than voice identity. All eight voices utilize the same underlying synthesis model, ensuring consistent computational requirements regardless of whether you select Aria or Atlas.

CPU usage variations

Since the model is uniform, CPU usage does not vary by voice. Processing is linear, with typical generation rates at approximately 25 characters per second for long texts; short phrases take a few seconds. Under load, requests queue, potentially increasing effective processing time but not altering per-request CPU intensity.

Latency differences

Latency is determined by queue position and text length, not voice selection. Short phrases take a few seconds to generate. Long texts process at real-time speed, so total latency equals generation time plus queue wait time. This applies uniformly across all voices.

  • Requests are processed one after another.
  • Queue wait time varies with load.
  • No parallel processing per voice.

File size impact

File size depends on the chosen response_format and duration, not the voice. MP3 files are smaller but compressed; WAV and FLAC offer higher fidelity at larger sizes. PCM provides raw data. Voice choice does not affect output byte count.

  • MP3: compressed, smaller size.
  • WAV/FLAC: lossless, larger size.
  • PCM: raw audio data.

Use Case Recommendations

Sound wave illustration for: Use Case Recommendations

The API supports audiobook narration by splitting long texts into 4096-character segments and joining the resulting audio files. This workflow ensures continuous playback for books, leveraging voices like Aria or Orion for consistent narrative tone.

Audiobook narration

Users can generate full-length books by sequentially processing segments. Developers must split the text into segments of up to 4096 characters and join the resulting audio files.

  • Ideal for LibriVox-style public domain content
  • Supports up to 4096 characters per request

Educational material reading

Educational platforms benefit from Sage’s calm instructional tone or Luna’s soft storytelling style. The API’s reliability ensures consistent audio generation for classroom or self-paced learning applications.

  • Clear pronunciation aids dyslexic readers
  • Speed adjustable from 0.25x to 4.0x

News content delivery

Atlas provides a neutral, news-like voice suitable for daily updates. Requests are processed one after another, so under load a request may wait in a queue.

  • Neutral tone suitable for professional contexts
  • Fast synthesis for timely updates

Limitations of Current Voices

Sound wave illustration for: Limitations of Current Voices

While AudioLex Access delivers consistent, high-fidelity synthesis for accessibility applications, developers should account for specific structural limitations in the current voice engine. Understanding these constraints ensures proper implementation for dyslexia-focused reading tools.

No emotion tags

The API does not support SSML or dynamic emotion injection. Voice delivery remains consistent without real-time tonal shifts for emphasis, which simplifies parsing but limits nuanced dramatic reading.

No custom cloning

Customers cannot upload audio to clone specific voices. The eight presets are fixed, ensuring uniform quality but preventing brand-specific voice customization without manual design work.

Fixed pitch ranges

Pitch is inherent to each preset and cannot be adjusted independently via parameters. Speed ranges from 0.25 to 4.0, but fundamental tone remains static per voice.

Limitations of Current Voices

Audio Format Impact on Quality

Sound wave illustration for: Audio Format Impact on Quality

The AudioLex Access API allows developers to select the output format that best balances bandwidth efficiency with audio fidelity for their accessibility applications. Understanding the technical trade-offs between compression and clarity is essential for optimizing both user experience and API costs.

MP3 compression effects

MP3 is one of the available formats, offering broad compatibility with minimal file size. For long-form reading applications, this significantly reduces bandwidth usage without noticeable degradation in speech intelligibility for most listeners.

  • Default format for most integrations
  • Optimal for bandwidth-constrained mobile devices
  • Negligible loss in speech clarity

WAV clarity benefits

WAV provides uncompressed, lossless audio, preserving every detail of the synthesized voice. This format is ideal for high-fidelity accessibility tools where users require the clearest possible audio signal, though it results in larger file sizes.

  • Lossless audio quality
  • Maximum clarity for sensitive listeners
  • Higher bandwidth consumption

Opus efficiency trade-offs

Opus delivers superior compression ratios compared to MP3, maintaining high quality at lower bitrates. This makes it an excellent choice for real-time reading features where latency and data usage are critical constraints.

  • High compression efficiency
  • Low latency suitable for streaming
  • Requires modern player support

User Preference Studies

Sound wave illustration for: User Preference Studies

Since the API lacks emotion tags, listeners rely on consistent pitch and steady pacing to maintain comprehension during extended reading sessions. This stability reduces cognitive load significantly.

Users report that predictable audio output allows for better focus on content decoding. The absence of variable emotional cues prevents distraction, making the synthetic speech suitable for long-form educational material where consistency outweighs expressive nuance.

Reader feedback analysis

Feedback consistently highlights the importance of clear enunciation over expressive variety. Users with dyslexia prioritize accurate phoneme delivery, which the current voice presets deliver without erratic intonation shifts.

The lack of emotion tags is generally accepted as a feature rather than a limitation, as it ensures uniform reading speed and tone throughout a document.

  • Clarity prioritized over expressiveness
  • Consistent pitch reduces cognitive load
  • No emotional variability distractions

Speed preference trends

Most users prefer speeds between 1.0 and 1.5x for complex text. The API supports speeds from 0.25 to 4.0, allowing fine-tuning. However, speeds above 2.0x may reduce comprehension for difficult material.

Slower speeds (0.25–0.75) are useful for initial exposure to new vocabulary. The real-time generation speed of roughly 25 characters per second provides a natural baseline for most readers.

  • Optimal speed: 1.0–1.5x
  • Support range: 0.25–4.0x
  • Real-time generation baseline

Voice familiarity impact

Familiarity develops over repeated use of the same voice preset.

Switching voices mid-document can cause minor disorientation. The eight distinct presets allow for variety, but consistency within a single reading session is recommended for optimal user experience.

  • Stick to one voice per session
  • Familiarity improves comfort
  • Switching may cause disorientation

Integration Complexity by Voice

Sound wave illustration for: Integration Complexity by Voice

The API maintains a uniform request structure across all eight voice presets. Developers send a single POST request to /v1/audio/speech with the voice parameter set to the desired preset. This consistency simplifies integration.

Parameter consistency

Each voice accepts the same standard parameters: model, input, response_format, speed, and optional language. The speed range (0.25–4.0) applies uniformly, allowing dyslexia-focused apps to adjust reading rates without recalibrating for specific voice characteristics. This predictability simplifies integration for indie developers building accessibility tools.

  • Uniform parameter set across all 8 voices
  • Standard speed range (0.25–4.0) applies to all
  • Optional language code works with any voice

Endpoint compatibility

All voices are served through the single POST https://api.6zzz.info/v1/audio/speech endpoint. There are no voice-specific endpoints or path variations. This reduces the number of network calls and simplifies error handling, as the same response schema applies regardless of the selected voice preset.

SDK support uniformity

Because the request shape mirrors the OpenAI speech endpoint, any official OpenAI SDK works by updating the base_url. This allows developers to use familiar client libraries without writing custom HTTP clients for each voice. No custom SDK is required, ensuring consistent behavior across all voice presets.

Cost Implications of Voice Choice

Sound wave illustration for: Cost Implications of Voice Choice

Selecting a voice for your accessibility application does not alter your operational costs. Every preset, from the warm narration of Aria to the neutral tone of Atlas, is processed by the same underlying engine and billed at the identical rate.

This uniformity eliminates the need to budget for premium tiers based on voice selection. Your financial planning remains straightforward, as the character count is the sole determinant of your expense, regardless of whether you generate content using Luna or Orion.

Uniform pricing per voice

All eight voices share the same base rate of $10 per 1 million characters. This means that switching from a female voice like Nova to a male voice like Milo incurs no additional cost. The engine complexity is consistent across all presets.

  • Single rate: $0.01 per 1,000 characters
  • No per-voice tiering
  • Identical engine for all 8 voices

No premium voice fees

Unlike platforms that charge extra for 'high-quality' or 'premium' voices, AudioLex Access applies a flat fee. Whether you use the soft storytelling of Luna or the energetic promo style of Felix, the charge per character remains exactly $0.00001.

Predictable billing

Your costs are directly proportional to usage. With failed requests refunded in full, you only pay for successful synthesis. Combined with the fixed rate, this allows indie developers to forecast expenses accurately without hidden fees for voice-specific attributes.

Future-Proofing Your Choice

Sound wave illustration for: Future-Proofing Your Choice

The API endpoint and request schema remain stable, ensuring your integration does not require frequent updates as underlying models evolve. This standardized structure allows indie developers to build once and maintain consistently for accessibility applications.

Switching voices is a simple parameter change in the voice field, requiring no code refactoring. This flexibility supports evolving user preferences without increasing technical debt or development time.

Standardized API structure

Our API follows a consistent JSON shape similar to industry standards, making it easy to adopt using existing HTTP clients or the OpenAI SDK with a simple base URL change. This reduces boilerplate code and accelerates initial deployment for developers focusing on user experience.

Ease of voice switching

Changing the voice profile requires only updating the voice parameter in your request body. This allows A/B testing of Aria, Nova, or Atlas with different user segments without altering your application logic or database schema.

Long-term support commitment

Your API key remains active until you issue a new one, ensuring continuity for your reading applications. There are no forced platform migrations or sudden infrastructure changes that would break existing integrations.

Listen to the 8 voices

Each sample below was generated by this API from the same sentence. WAV provides uncompressed, lossless audio, preserving every detail of the synthesized voice.

Avatar of the voice aria
ariafemale
Avatar of the voice nova
novafemale
Avatar of the voice luna
lunafemale
Avatar of the voice sage
sagefemale
Avatar of the voice felix
felixmale
Avatar of the voice milo
milomale
Avatar of the voice atlas
atlasmale
Avatar of the voice orion
orionmale

23 languages

Any voice speaks any of these languages.

Sources and further reading

  1. Wikipedia
  2. LibriVox public-domain audiobooks
  3. IETF RFC 7845: Ogg Opus
  4. Tether (USDT)
  5. OpenAI API reference: create speech
  6. LibriTTS-R corpus (OpenSLR 141)
  7. Creative Commons Attribution 4.0
  8. TRON developer hub: TRC-20
  9. MDN: Web Speech API
  10. W3C WAI: making audio and video accessible
  11. LibriTTS-R paper (arXiv 2305.18802)
  12. openai-python on GitHub

More on this site