ComfyUI Node
Maya1 TTS (AIO) Barebones
Maya1 TTS with standard widgets (for users with JavaScript rendering issues)
Maya1 TTS (AIO) Barebones
- audio
◄voice_descriptionRealistic male voice in the 30s age with american accent. Normal pitch, warm timbre, conversational pacing.►
◄textHello! This is Maya1 <laugh> the best open source voice AI model with emotions.►
◄model_name(No models folder found - see console for instructions)►
◄dtypebfloat16►
◄attention_mechanismsdpa►
◄devicecuda►
◄keep_model_in_vramtrue►
◄chunk_longformfalse►
◄max_new_tokens4000►
◄temperature0.40►
◄top_p0.90►
◄repetition_penalty1.10►
◄seed0►
Categoryaudio/maya1
Inputs (13)
| Name | Type | Default | Description |
|---|---|---|---|
| voice_description | STRING | Realistic male voice in the 30s age with american accent. Normal pitch, warm timbre, conversational pacing. | Describe your desired voice using natural language. Include: age (20s-50s), gender (male/female), accent (American/British/etc), pitch (low/normal/high), timbre (warm/gravelly/smooth), pacing (fast/conversational/slow), tone (happy/calm/energetic) |
| text | STRING | Hello! This is Maya1 <laugh> the best open source voice AI model with emotions. | Your script text to synthesize. Add emotion tags anywhere in the text (type manually - no visual buttons in barebones mode). All 17 available tags: <laugh>, <laugh_harder>, <giggle>, <chuckle>, <cry>, <sigh>, <gasp>, <whisper>, <angry>, <scream>, <snort>, <yawn>, <cough>, <sneeze>, <breathing>, <humming>, <throat_clearing> |
| model_name | COMBO | (No models folder found - see console for instructions) | Select Maya1 model from ComfyUI/models/maya1-TTS/ folder. Models are auto-discovered on startup. Download from: huggingface.co/maya-research/maya1 |
| dtype | COMBO | bfloat16 | Model precision. 4bit/8bit save VRAM but are SLOWER. Use float16/bfloat16 if you have 10GB+ VRAM for best speed. 4bit≈6GB, 8bit≈7GB, float16/bfloat16≈8-9GB, float32≈16GB |
| attention_mechanism | COMBO | sdpa | Attention algorithm. SDPA (default) is fastest for single TTS. Eager is standard PyTorch (slower). Flash Attention 2 helps with batch processing. Sage Attention is memory efficient |
| device | COMBO | cuda | Processing device. CUDA (GPU) is recommended for speed. CPU works but is much slower. Will auto-fallback to CPU if CUDA unavailable |
| keep_model_in_vram | BOOLEAN | true | Keep model loaded in VRAM after generation. True = faster repeated generations but uses VRAM. False = frees VRAM after each generation but slower |
| chunk_longform | BOOLEAN | false | Split long text into chunks at sentence boundaries with smooth crossfading. Enables unlimited audio length beyond the 18-20s limit. EXPERIMENTAL - may have quality/timing issues |
| max_new_tokens | INT | 4000100–16000 | Maximum NEW SNAC tokens to generate per chunk (excludes input prompt tokens). Higher = longer audio per chunk (~50 tokens/word). 4000 tokens ≈ 30-40s audio. Increase if audio cuts off too early |
| temperature | FLOAT | 0.400.1–2 | Controls randomness/creativity. Lower (0.1-0.3) = more consistent/predictable. Higher (0.5-1.0) = more varied/creative. 0.4 is official Maya1 recommendation |
| top_p | FLOAT | 0.900.1–1 | Nucleus sampling - controls diversity of token selection. 0.9 (default) works well for natural speech. Lower = more focused, higher = more diverse. Keep at 0.9 unless experimenting |
| repetition_penalty | FLOAT | 1.101–2 | Reduces repetitive speech patterns. 1.0 = no penalty, higher = stronger penalty against repetition. 1.1 is a good default. Increase to 1.2-1.3 if speech sounds too repetitive |
| seed | INT | 00–18446744073709550000 | Random seed for reproducibility. 0 = random seed each time. Set specific number (1-999999) for same result every time. Use control_after_generate widget to auto-increment/randomize |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| audio | AUDIO | — |