APIUp to 28% cheaper than official pricesTry the API →
Hermes
Back to News

Tiny-voca Packs Two AI Voices Into an 18 MB Model for $9

tiny-voca 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.

Tiny-voca Packs Two AI Voices Into an 18 MB Model for $9
Source
AlphaSignal
Published
Author
AlphaSignal
Read
2 min read

Takeaways

  • tiny-voca 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 .

Tiny-voca 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

ComponentProject specification
Parameters4.5 million, including waveform synthesis
VoicesF2 and M2
Training audioAbout 18 hours of synthetic English speech
Output24 kHz mono audio
Reported CPU speedAbout 70 times real time on four threads
Training run104,737 steps in 7.7 hours on one RTX 4090
ArtifactsPyTorch, ONNX, and Q4, Q8, and F16 GGUF files
Smallest published file8.7 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.

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Sources

  1. 01huggingface.co
  2. 02github.com
  3. 03github.com
  4. 04huggingface.co
  5. 05huggingface.co