Tiny-voca Packs Two AI Voices Into an 18 MB Model for $9 A 4.5-million-parameter English text-to-speech model called tiny-voca, released by its author n4ze3m on Hugging Face, ships two fixed voices (F2 and M2) in an 18 MB package and runs about 70 times faster than real time on four CPU threads. The model was trained from scratch on 17.6 hours of synthetic speech generated by Qwen3-TTS with no human recordings, and the entire project cost $9.56 in rented RTX 4090 time, including a final 7.7-hour, 104,737-step training run. On the Open TTS Leaderboard, tiny-voca's M2 voice ranks 8th and F2 ranks 13th among 50 English entries by average word error rate, with one entry placing above a 2-billion-parameter audio model; weights ship as PyTorch, ONNX and Q4/Q8/F16 GGUF files under Apache-2.0, with the smallest file at 8.7 MB. Tiny-voca Packs Two AI Voices Into an 18 MB Model for $9 A 4.5M parameter English TTS model trained only on synthetic speech runs 70x faster than real time on CPU and cost under $10 to build. - tiny-voca https://huggingface.co/n4ze3m/tiny-voca?utm source=alphasignal is a 4.5M parameter English TTS model, 18 MB, running 70x faster than real time on CPU. - Trained from scratch on 17.6 hours of synthetic speech generated by Qwen3-TTS, with zero human recordings. - Ranks 8th M2 and 13th F2 on the Open TTS Leaderboard out of 50 English models. - Entire project cost $9.56 in rented RTX 4090 time; final training run was 7.7 hours. - Uses a source-filter synthesizer 25k parameters instead of a neural vocoder for waveform generation. - Ships ONNX, PyTorch and GGUF weights under Apache-2.0; runs in-browser via tiny-voca-studio https://huggingface.co/spaces/n4ze3m/tiny-voca-studio?utm source=alphasignal . Tiny-voca https://huggingface.co/n4ze3m/tiny-voca?utm source=alphasignal is a new English text-to-speech model whose immediate training corpus contains only machine-generated audio. It ships two fixed voices in an 18 MB package. According to its author, generating the dataset, training the model, and running failed experiments consumed $9.56 in rented GPU time. That figure excludes development labor. Leaderboard gains, narrowly measured In the Open TTS Leaderboard results cited by the project, the M2 voice ranks eighth by average word error rate and F2 ranks thirteenth among 50 English entries. Both rows list CPU-only inference. One tiny-voca entry also sits above a 2 billion-parameter audio model. Word error rate compares an automated transcription of generated speech with the requested text, counting substitutions, insertions, and omissions. Lower scores indicate clearer pronunciation on the benchmark. Naturalness, expressiveness, speaker similarity, and latency require separate evaluation, so the ranking supports a focused claim about intelligibility. An 18 MB build | Component | Project specification | |---|---| | Parameters | 4.5 million, including waveform synthesis | | Voices | F2 and M2 | | Training audio | About 18 hours of synthetic English speech | | Output | 24 kHz mono audio | | Reported CPU speed | About 70 times real time on four threads | | Training run | 104,737 steps in 7.7 hours on one RTX 4090 | | Artifacts | PyTorch, ONNX, and Q4, Q8, and F16 GGUF files | | Smallest published file | 8.7 MB | How 4.5 million parameters speak 1. Text preparation: A phonemizer converts written English into phoneme IDs and stress markers. 2. Encoding and timing: A transformer encoder processes the phonemes, and a duration predictor assigns each phoneme a number of audio frames. This story is for Pro members You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.