📄 Development of an Amharic Text-to-Speech System Using Cepstral Method

Summary

Anberbir and Takara build AmhTTS, a text-to-speech system for Amharic, a widely spoken but poorly resourced Semitic language. They choose a parametric, rule-based design over concatenative synthesis because it needs less data, runs on small devices, and could be adapted to other Ethiopian Semitic languages. The speech inventory is a recording of every Amharic syllable, analysed into cepstral parameters. Synthesis uses a source-filter model with a Log Magnitude Approximation (LMA) filter. Amharic rhythm depends on syllable length rather than stress, so most prosodic work goes into duration rules. These cover geminated consonants, which the script does not mark and which the authors insert by hand, and "sixth order" syllables. A Fujisaki-style F0 contour is added over phrases found from punctuation. Native listeners recognised 98% of synthesized words, and sentences received a MOS of 3.2. The synthetic speech is clearly intelligible but only moderately natural. The paper is presented as the first rule-based Amharic TTS.

Key points

• Speech database and analysis: 196 syllables were recorded at 48 kHz and down-sampled to 10 kHz. They were analysed with 25.6 ms Hamming-windowed frames and a 10 ms shift. Each frame stores a voiced/unvoiced flag, a pitch period, and cepstral coefficients c[0]–c[29].
• Synthesis filter: the LMA filter is driven by an impulse train for voiced frames and white noise for unvoiced ones. Coefficient c[0] sets the gain.
• Prosody: this has three parts: gemination lengthening, sixth-order syllable duration control (with an optional epenthetic vowel /ix/), and a global intonation contour.
• Syllable joins: four connection rules depend on consonant type, using either smoothing or interpolation at boundaries. A 40 ms closure pause is inserted before plosives and glottal stops.
• Evaluation: four native listeners took part.
  • In the word test (200 common words), 40% of words contained geminates and 75% contained sixth-order syllables. Correct rates were 100% for original and resynthesized speech and 98% for rule-based speech.
  • In the sentence test, MOS was 5.0 for original, 4.7 for analysis/synthesis, and 3.2 for AmhTTS. Note that 3 is labelled "Fair" on the paper's own 1–5 scale, although the abstract calls 3.2 good.

Interesting ideas

• I1. When the script hides a phonologically meaningful feature (gemination separates alä "he said" from allä "there is"), the authors add an explicit manual mark rather than letting the system guess silently.
• I2. The evaluation includes three conditions: original speech, analysis/resynthesis, and full rule-based synthesis. This separates the loss caused by the parametric representation from the loss caused by the rules.
• I3. The first five of fifteen listening rounds were discarded as warm-up, so scores reflect judges who are used to the task rather than first-exposure noise.
• I4. How two stored units are joined is chosen by a small rule keyed on the type of the boundary segments (voiced vs. unvoiced, plosive), rather than one generic join.
• I5. Prosodic phrases are approximated cheaply from punctuation, which is crude but transparent and good enough to drive an intonation model.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): ambiguities that the surface form underdetermines could be carried as an explicit, visible feature in morphology. Downstream stages and explanations would then show where a guess was made.
• I2 (learn): we could run frames and answers from gold trees, from reparsed gold trees, and from fully automatic trees. Each stage's share of the error would then be measured rather than guessed.
• I3 (learn): when human or LLM judges rate generated answers or paraphrases, a calibration batch whose scores are thrown away could stabilise the ratings.
• I4 (back): when realising text from the meaning graph, the choice of connectives, pronouns or punctuation between clauses could be selected by boundary-type rules. This would keep each choice explainable.
• I5 (back): punctuation-based chunking could give a simple, inspectable baseline for sentence-level planning in generation before anything learned is added.