huggingface.co web signal

Nazeem releases tiny-voca, 4.5M-param English TTS for $9.56

TL;DR

  • tiny-voca is a 4.5-million-parameter English TTS trained on one RTX 4090 in 7.7 hours for $9.56 of GPU time.
  • The checkpoint runs ~70x faster than real time on four CPU threads and ships as PyTorch, ONNX, and GGUF down to 8.7 MB.
  • Its M2 voice ranks 8th and F2 ranks 13th on the Open TTS Leaderboard v1.6.3 with sub-0.8% average WER, under Apache-2.0.

Muhammed Nazeem has published tiny-voca, a 4.5-million-parameter English text-to-speech model trained on a single RTX 4090 in 7.7 hours for $9.56 of GPU time. The model card lists 104,737 training steps over 17.6 hours of synthetic speech, and reports the checkpoint runs "~70x faster than real time on four CPU threads."

On the Open TTS Leaderboard snapshot of November 16, 2026 (v1.6.3), the M2 male voice sits at rank 8 with a 0.733% average word error rate and a UTMOS22 of 4.34; the F2 female voice holds rank 13 with 0.773% WER and UTMOS22 4.43, in a field the card describes as "competing against 50 full-coverage English models."

The release ships as PyTorch weights (18.2 MB), an ONNX export (21.4 MB), and GGUF quantizations down to 8.7 MB for the 4-bit variant, under Apache-2.0. The ONNX build runs inside a browser tab via onnxruntime-web, capped at 512 tokens and 32 seconds per utterance.

Nazeem is direct about what the model is not. The card lists its limitations as "English only, two fixed voices, no voice cloning," notes that "identical text always reads the same way" because there is no prosody sampling, and says "training data is synthetic, limiting naturalness."