TextSynth

TextSynth provides access to multiple large language models through their REST API, including capabilities for text-to-image generation, text-to-speech synthesis, and speech-to-text transcription. The company's models, developed from 2020 onwards, include Mistral, Llama, Stable Diffusion, and Whisper architectures, with parameter counts ranging from 6 billion to 70 billion. This technical overview examines TextSynth's capabilities, API fundamentals, text generation services, content creation tools, and pricing structure.

Company Overview

TextSynth provides access to multiple large language models through their REST API and includes tools for text-to-image generation, text-to-speech synthesis, and speech-to-text transcription. The company's models, developed from 2020 onwards, include Mistral, Llama, Stable Diffusion, and Whisper architectures, with parameter counts ranging from 6 billion to 70 billion.

The API supports various text processing capabilities, including grammar-constrained generation and JSON schema-driven text completion. Response outputs include the generated text, metadata about prompt handling, and detailed token information. Text generation requests can include constraints through BNF grammar syntax, supporting complex formatting rules and structured output.

For content creation, TextSynth offers direct text-to-image generation with stable_diffusion models and automated speech-to-text transcription using the whisper_large_v3 engine. The text-to-image service requires prompt text and supports customization through parameters like guidance scale and seed value. Speech-to-text functionality accepts multiple audio formats, automatically detects language when provided, and returns detailed segment-level transcription data.

The company implements custom inference optimizations to reduce costs while operating all servers in France. Their pricing structure varies by model complexity, with base costs of $0.20 per million input tokens across most language models. Image generation pricing stands at $0.005000 per minute, while text-to-speech synthesis incurs costs of €0.0001 per character for standard service and €0.00005 for real-time requests. All transactions utilize a credit-based system, managed through Stripe payments with balances valid for one year.

API Fundamentals

TextSynth's API operates with a straightforward request-response model. To generate text, users must provide either a grammar definition following Backus-Naur Form (BNF) syntax or a JSON schema, but not both. The API processes requests by returning several key pieces of information in its response:

• The completed text

• A boolean indicating if it's the last answer

• A boolean indicating if the prompt was truncated

• The reason for completion (stopped at a certain point, reached maximum length, or violated grammar rules)

• The number of input and output tokens generated

• The format of the output text, typically returned as a string or an array of strings

The BNF grammar syntax allows for sophisticated text generation constraints through production rules that define how non-terminal symbols can be replaced by other non-terminal symbols or terminal strings. These rules support basic data types, nested structures, and repetition patterns while prohibiting left recursion to maintain computational efficiency.

For structured data processing, TextSynth supports JSON Schema syntax with constraints for various data types including objects, arrays, strings, integers, numbers, booleans, and null values. The schema requires at least one required property for objects and supports array minimum length constraints and string enumeration options. The company's models process inputs with 8K context length, allowing for moderate-sized prompt contexts while maintaining efficient memory usage.

Text Generation Services

TextSynth's text generation capabilities span multiple domains through their REST API, including text completion, chat functions, and translation services.

Text completion is facilitated through their API endpoints, which support both grammar-constrained generation via Backus-Naur Form (BNF) syntax and structured data processing with JSON Schema. The company's custom inference optimizations enable efficient processing on standard GPUs and CPUs, reducing costs for users.

For chat applications, TextSynth automatically formats prompts according to the selected model's preferred chat template, allowing for seamless integration with conversational interfaces. The translation endpoint supports multiple engines, including the MADLAD400 model, which enables translation between 400+ languages through its multilingual capabilities.

The API structure for these services mirrors their text completion endpoint, using a POST request to the relevant engine ID endpoint. For image generation, users interact with the stable_diffusion model through a separate text_to_image endpoint, specifying parameters like guidance scale and seed value for controlled output. Speech-to-text functionality leverages the whisper_large_v3 engine, accepting multiple audio file formats and returning detailed segment-level transcription data.

To facilitate development, TextSynth provides comprehensive API documentation covering endpoint structure, request parameters, and response formats. The company's pricing model distinguishes between input and generated tokens, with higher-parameter models commanding higher rates for token processing. All transactions occur through Stripe, utilizing a credit-based system valid for one year, with minimum purchase requirements of €20 or $20.

Content Creation Tools

TextSynth offers specialized content creation tools through their REST API, including text-to-image generation and speech-to-text transcription capabilities. The text-to-image service generates visual content from textual prompts using the stable_diffusion model, while speech-to-text functionality transcribes audio files into written text (Context: Speech-to-Text Transcription).

The text-to-image generation process requires prompts specifying the desired image content, with optimal results achieved using 512x512 pixel images (Context: Text-to-Image Generation). Users can customize the generation process through parameters including guidance scale, seed value, and image strength (Context: Text-to-Image Generation). The service returns generated images in base64-encoded JPEG format, allowing straightforward integration into various applications (Context: Text-to-Image Generation).

For speech-to-text transcription, TextSynth's whisper_large_v3 engine processes audio files up to 2 hours in length across multiple formats (Context: Speech-to-Text Transcription). The system automatically detects languages when provided with audio, returning detailed segment-level transcription data formatted as JSON (Context: Speech-to-Text Transcription). Implementation requires including authorization headers and specifying audio file parameters in API requests (Context: API Implementation).

These content creation tools support diverse applications through well-documented API endpoints and custom inference optimizations that enhance processing efficiency on standard hardware (Context: API Fundamentals). The company's comprehensive pricing structure distinguishes between input and generated token processing costs across their services (Context: Pricing Information). Full documentation and implementation examples are available to assist developers in integrating these tools into their applications (Context: API Documentation).

Pricing and Usage

The basic TextSynth service is free but rate-limited, with users restricted to 200 generated tokens per request and subject to regular captcha requirements. To enable unlimited usage, users can purchase credits through the company's payment system, which accepts minimum purchases of €20 or $20 via Stripe. Credit balances remain valid for one year, with unused credits expiring at the end of this period.

The company employs a token-based pricing model, where each token represents approximately four English characters of information processed by their language models. The pricing structure varies significantly based on the complexity of the model in use, ranging from $0.20 per million input tokens down to €0.0001 per second for real-time speech-to-text processing.

For the company's language generation models, ranging from 6 billion to 70 billion parameters, the cost structure differs between input and generated tokens. Basic models like GPT-J charge $0.20 per million input tokens and $2.00 per million generated tokens, while larger models such as Llama3.3 70B instruct command $0.70 per million input tokens and $7.00 per million generated tokens.

Text-to-image generation utilizes the stable_diffusion model and operates on a minute-based pricing structure, charging $0.005000 per minute. Speech-to-text transcription employs the whisper_large_v3 engine and offers two service levels: standard at €0.003000 per minute and real-time at €0.009000 per minute.

All API transactions utilize TextSynth's built-in credit system, with pricing calculations based on both input and generated tokens. The company's custom inference optimizations enable efficient processing on standard GPUs and CPUs, maintaining lower costs compared to alternative implementations.