# grok-text-to-speech **Model:** `xai/grok-text-to-speech` | **Provider:** Replicate | **Creator:** xAI | **Categories:** Text-to-Speech, Audio --- ## Overview **grok-text-to-speech** is xAI's production-grade voice synthesis model that converts written text into natural, expressive speech audio. Designed for developers, content creators, and product teams, it delivers high-fidelity audio across five distinct voices, twenty languages, and multiple output formats — from broadcast-quality MP3 to telephony-optimized μ-law encoding. On CRAISEE, this model is the go-to solution for any workflow that requires reliable, expressive, and linguistically versatile text-to-speech conversion at scale. --- ## Key Capabilities - **Five distinct voices** — `eve` (energetic, upbeat), `ara` (warm, friendly), `rex` (confident, clear), `sal` (smooth, balanced), and `leo` (authoritative, strong), each with a unique personality suited to different content contexts. - **Twenty-language support** with automatic language detection, eliminating the need to manually specify language in multilingual workflows. - **Inline speech tags** for granular delivery control — insert `[pause]`, `[long-pause]`, `[laugh]`, `[sigh]`, and `[breath]` at precise points in your script. - **Wrapping speech tags** to modify entire spans of text — `...`, `...`, and `...` for nuanced emotional and pacing control. - **Multiple audio output formats** — MP3 (default), WAV, raw PCM, μ-law, and A-law, covering use cases from podcast production to telephony IVR systems. - **Configurable audio quality** — sample rates from 8 kHz to 48 kHz, and MP3 bit rates from 32 to 192 kbps. - **Long-form input support** — accepts up to 15,000 characters per request, making it viable for narrating full articles, chapters, or scripts in a single call. - **Automatic text normalization** — numbers, currencies, dates, and abbreviations are spoken naturally without manual preprocessing. - **High run volume** — with over 1.9 million runs logged, this model is proven at production scale. --- ## Best Use Cases Choose **grok-text-to-speech** on CRAISEE when your project requires any of the following: ### Content Narration & Podcasting The model's long-form input capacity (15,000 characters) and high-fidelity MP3 output at up to 192 kbps make it ideal for narrating blog posts, articles, newsletters, and podcast scripts. Use `rex` or `leo` for authoritative documentary-style narration, or `ara` for conversational podcast formats. ### E-Learning & Instructional Audio `leo`'s authoritative, commanding tone is purpose-built for instructional content. Combined with `...` wrapping tags to emphasize key concepts and `[pause]` tags to give learners processing time, this model produces professional-grade e-learning audio without a recording studio. ### Multilingual Product Localization With 20 languages and automatic detection, grok-text-to-speech is the strongest choice on CRAISEE for teams building multilingual apps, localized marketing content, or international customer-facing audio. A single API call handles language switching automatically. ### Telephony & IVR Systems The availability of μ-law and A-law output formats at 8 kHz sample rates makes this model directly compatible with telephony infrastructure. It is the only model on CRAISEE optimized for telephony-grade audio encoding out of the box. ### Interactive Storytelling & Games Speech tags like `[laugh]`, `[sigh]`, and `...` enable emotionally expressive character dialogue. Writers and game developers can script nuanced vocal performances without voice actors. ### Accessibility Features Developers building screen readers, read-aloud features, or assistive tools benefit from the model's natural prosody, text normalization, and reliable long-form handling. --- ## How to Use on CRAISEE Follow these steps to run grok-text-to-speech on the CRAISEE platform: **Step 1 — Navigate to the Model** Search for `grok-text-to-speech` in the CRAISEE model library, or browse under the **Text-to-Speech** category. Select the model card for `xai/grok-text-to-speech`. **Step 2 — Open the Playground or API Panel** CRAISEE provides a no-code Playground for quick testing and a full API panel for integration. Start with the Playground to preview voices and speech tag behavior before committing to an integration. **Step 3 — Enter Your Text** Paste or type your script into the text input field. You may include inline speech tags directly in the text (e.g., `Hello [pause] and welcome to our platform.`). The model accepts up to 15,000 characters per request. **Step 4 — Select Your Voice** Choose from `eve`, `ara`, `rex`, `sal`, or `leo` in the voice selector. Refer to the voice personality descriptions to match tone to your content type. **Step 5 — Configure Output Settings** Set your preferred output format (MP3, WAV, PCM, μ-law, or A-law), sample rate (8–48 kHz), and bit rate (32–192 kbps for MP3). For most content creation use cases, the defaults (MP3, 44.1 kHz, 128 kbps) are appropriate. **Step 6 — Set Language (Optional)** If your text is in a specific language, select it from the language dropdown. Leave it on auto-detect for multilingual or mixed-language inputs. **Step 7 — Run the Model** Click **Run** in the Playground or submit your API request. The model returns an audio file stream that you can play back directly in CRAISEE or download to your local environment. **Step 8 — Iterate** Adjust speech tags, voice selection, or pacing parameters and re-run until the output matches your creative intent. CRAISEE saves your run history for easy comparison. --- ## Input Parameters The following parameters control model behavior. All are available in both the CRAISEE Playground and via the API. | Parameter | Type | Default | Description | |---|---|---|---| | `text` | `string` | *(required)* | The input text to synthesize. Supports inline and wrapping speech tags. Maximum 15,000 characters. | | `voice` | `string` | `eve` | Voice selection. Options: `eve`, `ara`, `rex`, `sal`, `leo`. | | `language` | `string` | `auto` | Target language code (e.g., `en`, `ko`, `fr`, `de`). Set to `auto` for automatic detection. | | `output_format` | `string` | `mp3` | Audio encoding format. Options: `mp3`, `wav`, `pcm`, `ulaw` (μ-law), `alaw` (A-law). | | `sample_rate` | `integer` | `44100` | Audio sample rate in Hz. Range: 8,000–48,000. Use 8,000 for telephony; 44,100 or 48,000 for high-fidelity output. | | `bit_rate` | `integer` | `128` | MP3 bit rate in kbps. Range: 32–192. Only applies when `output_format` is `mp3`. | ### Speech Tag Reference **Inline tags** (inserted at a specific point in text): | Tag | Effect | |---|---| | `[pause]` | Short pause in speech delivery. | | `[long-pause]` | Extended pause, useful for dramatic effect or section breaks. | | `[laugh]` | Inserts a natural-sounding laugh. | | `[sigh]` | Inserts an audible sigh. | | `[breath]` | Inserts an audible breath, adding realism to long passages. | **Wrapping tags** (applied to a span of text): | Tag | Effect | |---|---| | `...` | Delivers the enclosed text in a whispered tone. | | `...` | Reduces speaking pace for the enclosed text. | | `...` | Lowers volume and softens delivery for the enclosed text. | --- ## Tips & Best Practices **1. Match Voice to Content Tone Before Scripting** The single highest-impact decision in a TTS workflow is voice selection. Audition all five voices with a representative sample of your actual content — not placeholder text — before finalizing your choice. `leo` underperforms in casual conversational content just as `eve` can feel mismatched in formal legal narration. Correct voice selection eliminates the need for compensatory speech tag overuse. **2. Use Speech Tags Sparingly and Strategically** Speech tags are most effective when used to mark genuinely meaningful moments — a dramatic `[long-pause]` before a key reveal, a `` for an aside, a `[breath]` to break up a long technical passage. Overloading a script with tags produces unnatural, choppy audio. The model's prosody is strong enough to handle natural pacing without heavy tagging in most cases. **3. Leverage Text Normalization — Don't Fight It** grok-text-to-speech automatically normalizes numbers, currencies, dates, and abbreviations. Write `$1,500` rather than `one thousand five hundred dollars`, and `Dr.` rather than `Doctor` — the model handles the spoken form correctly. Manually spelling out normalized forms can actually degrade naturalness by disrupting the model's internal text analysis. **4. Optimize Sample Rate and Bit Rate for Your Delivery Channel** For web streaming or podcast distribution, 44.1 kHz at 128 kbps MP3 is the optimal balance of quality and file size. For telephony or IVR, switch to `ulaw` format at 8 kHz — this is the standard encoding for phone systems and will integrate directly without transcoding. For archival or professional post-production, use WAV at 48 kHz. Mismatching format to channel is the most common technical error in TTS deployments. **5. Batch Long Documents Into Logical Segments** Although the model supports up to 15,000 characters per request, very long inputs can produce subtle prosody drift across the audio. For documents longer than approximately 3,000–5,000 characters, split at natural section boundaries (chapter breaks, heading transitions) and concatenate the resulting audio files. This also gives you per-section voice or pacing control, and makes re-generation of edited sections faster and cheaper. --- ## Comparison ### grok-text-to-speech vs. Other TTS Models on CRAISEE **grok-text-to-speech is the definitive choice on CRAISEE for production-scale, multilingual, expressive text-to-speech.** Its combination of long-form input support, telephony format compatibility, and inline speech tag control places it in a distinct category from most alternative TTS models available on the platform. | Feature | grok-text-to-speech | Typical Alternative TTS Models | |---|---|---| | Max input length | 15,000 characters | Often 1,000–5,000 characters | | Voice options | 5 distinct personalities | Varies (often 2–4) | | Language support | 20 languages, auto-detect | Often fewer, manual selection required | | Speech tag control | Inline + wrapping tags | Rarely supported | | Telephony formats (μ-law, A-law) | ✅ Native support | ❌ Usually absent | | Text normalization | ✅ Automatic | Often manual preprocessing required | | Proven run volume | 1.9M+ runs | Varies | **Where alternatives may be preferred:** - Models fine-tuned for a specific single language may produce marginally more native-sounding output in that language compared to a multilingual model. - If your use case requires voice cloning or custom voice training from audio samples, grok-text-to-speech does not support this — a dedicated voice cloning model would be more appropriate. - For ultra-low-latency streaming TTS in real-time conversational applications, evaluate whether the model's response characteristics meet your latency requirements against specialized streaming TTS solutions. **Bottom line:** For the vast majority of content creation, localization, accessibility, and telephony use cases on CRAISEE, **grok-text-to-speech delivers the broadest feature coverage, the most flexible output options, and the most proven reliability of any TTS model available on the platform.** --- *Last updated for CRAISEE model library. Model slug: `xai/grok-text-to-speech`. Provider: Replicate. Run count at time of writing: 1,934,168+.*

grok-text-to-speech — AI Model Guide