​
Voicelab Models

Voicelab provides access to state-of-the-art open-source text-to-speech models.

​
Sesame CSM-1B

State-of-the-art conversational speech model

This model is optimized and certain voices are professionally cloned. This means that we offer fast, low-cost inference for this model, and that select voices will have more consistent quality.

  • Parameters: 1 billion
  • Specialization: Natural conversation and dialogue
  • Languages: English
  • Key Features:
    • Exceptional naturalness in conversational contexts
    • Low latency inference
    • Consistent voice quality across long generations
    • Optimized for real-time applications

​
Dia

general-purpose TTS model

This model is not yet optimized or post-trained. Generation quality may be inconsistent.

  • Parameters: 1.6 billion
  • Specialization: General-purpose text-to-speech
  • Languages: English
  • Key Features:
    • Accepts emotive tokens (e.g. laughing, sighing)
    • Efficient inference
    • Less stable than CSM-1B

Emotive Tags: (laughs), (clears throat), (sighs), (gasps), (coughs), (singing), (sings), (mumbles), (beep), (groans), (sniffs), (claps), (screams), (inhales), (exhales), (applause), (burps), (humming), (sneezes), (chuckle), (whistles)

​
Orpheus

Expressive and emotional speech synthesis

This model is not post-trained; the hosted weights are as-is.

  • Parameters: 750 million
  • Specialization: Emotional and expressive speech
  • Languages: English (primary)
  • Key Features:
    • Advanced emotional control
    • Dynamic prosody and intonation
    • Character voice capabilities
    • Rich vocal expressions

Emotive tags: <laugh>, <chuckle>, <sigh>, <cough>, <sniffle>, <groan>, <yawn>, <gasp>

​
Coming Soon

Kokoro, Chatterbox, Kyutai TTS

​
Model Specifications

​
Supported Features

FeatureSesame CSM-1BDiaOrpheus
Voice Cloning✅✅✅
Emotive Tokens❌✅✅
Multi-speaker✅✅✅
Real-time Streaming✅✅✅
Custom Fine-tuning✅✅✅

While CSM-1B doesn’t accept emotive tokens (laugh/sigh/etc.), it has the capacity to generate these artifacts in output audio based on conversational context.