Original title: n4ze3m/tiny-voca
An English TTS model uploaded by n4ze3m on 11 14, which the author states has a total of 4,537,291 parameters including the waveform generation component, outputs 24 kHz mono audio, and offers two American English voices: F2 (female) and M2 (male). It uses a source-filter synthesizer instead of a neural vocoder. The training data consists of approximately 17.6 hours, all synthesized by Qwen3-TTS with no human recordings. PyTorch weights and an ONNX version compatible with onnxruntime-web are provided.
Why it matters The entire model has only about four to five million parameters and does not require a GPU, making it suitable for adding English text-to-speech capabilities to offline applications, web pages, or embedded devices
Upvotes 77 · Likes 77 · Downloads 40
Reasons
- The author states that on 4 CPU threads it runs at approximately 70 times real-time speed, and the model card notes it can run on mobile devices or in browsers
- The dataset, training code, and phoneme conversion tool dizione are all public; the author states the entire project cost 9.56 dollars on a single rented RTX 4090
- The model card provides 4 example audio clips (one set for each voice), a complete parameter table, and limitations
Usage
- pip install git+https://github.com/n4ze3m/dizione git+https://github.com/n4ze3m/tiny-voca,下载模型后设 DIZIONE_DATA=tiny-voca/phonemizer, then use python -m tiny_voca.infer --checkpoint tiny-voca/tiny-voca.pt --voice F2 --text "…" --output hello.wav
- Adjustable parameters include --speed (speech rate), --pitch-shift (0.058 approximately one semitone), and --pitch-range; long texts are split by sentence and then concatenated
- The ONNX version supports a maximum of 512 phonemes and 32 seconds per run, requiring prior conversion to phonemes using dizione; output occasionally exceeds 1.0, so scale according to the model card example before saving as 16-bit audio
Risks
- Supports only English and two fixed voices; cannot clone voices or support streaming; the same sentence is read identically each time, and foreign names and uncommon spellings are often mispronounced
- The author states the voice is clear but has a slight synthetic quality; F2 sounds younger than designed. Since the training data itself is synthesized speech, naturalness will not exceed that of the generating Qwen3-TTS
- Evaluation consists only of Whisper word error rates for 8–16 sentences; the author clarifies this is a self-check and no formal listening tests were conducted. This entry has not been run or auditioned