--- license: apache-2.0 language: - en pipeline_tag: text-to-speech tags: - text-to-speech - onnx datasets: - n4ze3m/tiny-voca base_model: - n4ze3m/tiny-voca --- # **tiny-voca-GGUF** > **tiny-voca** is an ultra-lightweight, open-source (Apache-2.0) English text-to-speech (TTS) model created by Muhammed Nazeem (`n4ze3m`) that packs its entire architecture—including waveform generation—into just **4.54 million parameters** (~18.2 MB in PyTorch or ~21.4 MB in ONNX). Trained from scratch in under 8 hours on a single RTX 4090 (costing $9.56 total) using 17.6 hours of synthetic speech generated by Qwen3-TTS, it converts text to phonemes via the custom `dizione` phonemizer and uses transformer encoders paired with a fast **source-filter synthesizer** rather than a heavy neural vocoder to output 24 kHz mono audio. Supporting two fixed American English voices (**F2** female and **M2** male) with adjustable speed, pitch shift, and pitch range, it runs roughly **70x faster than real time on four CPU threads**—making it ideal for client-side browser or mobile deployment—and achieves competitive intelligibility (ranking 8th and 13th on the November 2026 Open TTS Leaderboard with sub-0.8% average Word Error Rate), though it lacks streaming, voice cloning, and prosody variation. > [!NOTE] > Use F32 for the best quality. Supports `llama-tts` via llama.cpp. ## Model Files | File Name | Quant Type | File Size | File Link | Description | |-----------|------------|-----------|-----------|-------------| | tiny-voca.F32.gguf | F32 | 6.42 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.F32.gguf) | Full 32-bit float weights. Maximum precision, largest file size. | | tiny-voca.BF16.gguf | BF16 | 3.28 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.BF16.gguf) | Full BF16 weights. High quality, excellent balance of precision and size. | | tiny-voca.Q3_K_L.gguf | Q3_K_L | 1.12 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.Q3_K_L.gguf) | Lower quality but usable, good for very low RAM availability. | | tiny-voca.Q3_K_M.gguf | Q3_K_M | 1.07 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.Q3_K_M.gguf) | Low quality, smallest file size. | | tiny-voca.Q4_K_M.gguf | Q4_K_M | 1.29 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.Q4_K_M.gguf) | Good quality, default size for most use cases, *recommended*. | | tiny-voca.Q4_K_S.gguf | Q4_K_S | 1.18 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.Q4_K_S.gguf) | Slightly lower quality with more space savings, *recommended*. | | tiny-voca.Q5_K_M.gguf | Q5_K_M | 1.37 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.Q5_K_M.gguf) | High quality, *recommended*. | | tiny-voca.Q5_K_S.gguf | Q5_K_S | 1.27 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.Q5_K_S.gguf) | High quality with slight space savings, *recommended*. | | tiny-voca.Q6_K.gguf | Q6_K | 1.76 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.Q6_K.gguf) | Very high quality, near perfect, *recommended*. | | tiny-voca.Q8_0.gguf | Q8_0 | 1.8 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.Q8_0.gguf) | Extremely high quality, generally unneeded but max available standard quant. | | tiny-voca.mmproj-bf16.gguf | mmproj-bf16 | 8.14 MB | [Link](https://huggingface.co/prithivMLmods/tiny-voca-GGUF/blob/main/tiny-voca.mmproj-bf16.gguf) | Multimodal projection file in BF16 format. Used for audio/voice capabilities. | ## llama.cpp LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp Quantized using llama.cpp by [prithivMLmods](https://huggingface.co/prithivMLmods).