Trelis Voice Router
Trelis Router gives you one spend-limited key for accessing transcription (ASR) and text to speech models. It’s like OpenRouter, but for voice. You put your api keys for voice providers into Trelis Router, and Router gives you back spend limited keys.
I created the tool because we use it internally at Trelis for evaluating frontier voice models easily. Take a look at router.trelis.com, cheers, Ronan
Cheers, Ronan
🗝️ Trelis All Access (7 Github Repos, Support via Github Issues & Private Discord)
🖥️ Data Prep + Training via UI or with Claude Code: Try it out: studio.trelis.com
🤝 White Glove Data Preparation + Training Services: Learn More
Timestamps:
0:00 Introduction to Trelis Router
0:06 Adding provider API keys at router.Trelis.com
1:07 Browser-based text-to-speech demo with 11 Labs
1:35 Pricing: 100 free requests monthly, then 5% charge
2:10 Integration with Trelis Studio for model evaluation
2:45 Comparing proprietary vs open source speech-to-text models
3:16 Use cases: key management and spend-limiting for agents
Trelis Voice Router: A Unified API for Voice Models
Trelis Voice Router provides a single API endpoint for accessing multiple voice models across transcription and text-to-speech services. The service operates on a bring-your-own-key model, allowing users to connect their existing provider credentials through one interface.
How the Service Works
Users access the platform at router.Trelis.com and add API keys for their chosen providers through the API keys section. Once provider keys are configured, users generate a spend-limited Trelis key that works across all connected services. This single key replaces the need to manage multiple provider credentials in applications.
The service accepts requests for models not currently supported through a suggestions system in the bugs section.
API Features
The Trelis Router API offers both synchronous and asynchronous endpoints. The platform can accept multiple requests simultaneously and batches them before forwarding to providers. Both speech-to-text and text-to-speech operations are supported, though text-to-speech currently has limited model availability.
A browser-based testing interface allows users to verify model functionality before implementing API calls. For example, text-to-speech synthesis can be tested with providers like 11 Labs directly through the web interface.
Pricing Structure
The first 100 requests each month incur no charges from Trelis beyond the costs from the underlying provider keys. After 100 requests, Trelis adds a 5% fee on top of provider costs. This pricing model matches Open Router’s approach.
Currently, users cannot spend Trelis credits directly - all charges route through the provider keys users have configured. This differs from some competing services but follows the early operational model used by Open Router.
Integration with Trelis Studio
Trelis Router API keys work within Trelis Studio, a model evaluation platform. This integration enables comparison testing between proprietary and open source voice models.
On the evaluation page, users can select models from providers like 11 Labs (such as Scribe V2) and compare them against open source alternatives including Qwen models and Mistral’s VoxTrial model. Users select a dataset from sources like Hugging Face, specify which split to evaluate (such as validation), and run the comparison.
Evaluation Metrics
The Studio platform calculates word error rate and character error rate for transcription models. Results display after processing completes. For example, a test run on a six-sample dataset returned a 17% word error rate and 10% character error rate.
This evaluation capability supports workflows where teams train base open source models, test performance improvements, and benchmark against proprietary alternatives.
Use Cases
The platform serves several practical purposes:
Model Development: Teams training custom models for clients use the evaluation features to measure performance against commercial alternatives. This applies to fine-tuning services where baseline comparisons determine whether custom training provides sufficient value.
Developer Access Management: Organizations can provide multiple team members access to various models through a single API key rather than distributing separate credentials for each provider. This simplifies configuration management when keys need to be updated or rotated.
Cost Control: The spend-limiting feature allows API keys to be capped at specific amounts. This proves useful when keys are passed to automated agents or systems where uncontrolled spending could occur.
Multi-Model Testing: During evaluation phases, teams can test multiple providers without reconfiguring applications for each service’s specific API format.
Technical Implementation
The service acts as a routing layer between applications and voice model providers. When a request arrives with a Trelis Router key, the system:
Identifies which provider key to use based on the requested model
Translates the request to the provider’s API format
Forwards the request to the appropriate service
Returns the response to the requesting application
For asynchronous operations, the platform queues requests and manages batching to optimize provider API usage.
Model Availability
Available models are listed on a dedicated models page with code examples for both synchronous and asynchronous implementations. The current model list includes various speech-to-text options, with text-to-speech having more limited availability. Users can request additional model support through the platform’s feedback system.
Summary
Trelis Voice Router consolidates access to multiple voice AI providers through a single API key with spend controls. The bring-your-own-key approach means users maintain direct relationships with providers while gaining simplified access management. Integration with Trelis Studio adds evaluation capabilities for teams comparing model performance across proprietary and open source options. The service charges nothing for the first 100 monthly requests, then adds 5% beyond provider costs.