tiny-voca-fr

A 4.5M-parameter standard French text-to-speech model with two voices. This page runs it in your browser: nothing you type leaves your device.

Loading…
4,532,683parameters, waveform generator included
21 MBmodel file (ONNX, float32)
24 kHzmono output
2 voicesFR_F1 and FR_M1, standard French
14.1 hoursof synthetic training speech
Apache-2.0code and weights
  1. Voice

    FR_M1 has a faint mechanical buzz

    On most sentences the male voice carries a low buzz in the background. It stays fully intelligible, but sounds less natural than FR_F1, which has no such artefact. At this low pitch (about 95 Hz) the model's voiced sound comes out more regular than the teacher's.

  2. Phonemes

    Some liaisons are missed

    eSpeak NG skips some required liaisons: comment allez is read without the [t].

  3. Phonemes

    A few false liaisons

    The opposite also happens: eSpeak NG links words that should stay apart, as in Dupont et.

  4. Phonemes

    fils (son) is read as [fil]

    eSpeak NG gives the pronunciation of fils (threads) instead of [fis]. Sentences containing the word were left out of the training data.

  5. Prosody

    A little fast, with short pauses

    Speech runs about 8% faster than the teacher recordings, and pauses between phrases are shorter.

    Workaround: lower Speed to 0.90–0.95.