Amharic Text-to-Speech: Ejectives, Vowel Orders and Neural Architectures

1. Glottalic Ejectives and Vowel Orders

Amharic has a set of non-pulmonic consonants known as ejectives:

Ejective Series:
  ጥ (tʼ)   - Alveolar ejective stop
  ቅ (kʼ)   - Velar ejective stop
  ጭ (tʃʼ)  - Post-alveolar ejective affricate
  ጽ/ፅ (tsʼ)- Alveolar ejective affricate
  ጵ (pʼ)   - Bilabial ejective stop
  ቋ (kʼʷ)  - Labialized velar ejective stop

How Ejectives Are Produced

Pulmonic stops (English /p/, /t/, /k/) use air from the lungs. Ejectives use a glottalic airstream:

1. The vocal folds seal the glottis.
2. The larynx rises, compressing the air trapped behind the oral constriction.
3. The oral closure releases in a sharp burst.
4. The glottis opens, starting voicing for the following vowel.

A model that misses the glottalic closure flattens ejectives into English-style aspirated stops ([tʰ], [kʰ]).

The Seven Vowel Orders

Order	Traditional Name	IPA Phonetic Value	Acoustic Description
1st	Gə’əz	[ɐ] or [ɜ]	Mid-central open-mid vowel
2nd	Ka’əb	[u]	High-back rounded vowel
3rd	Saləs	[i]	High-front unrounded vowel
4th	Rabə	[a]	Low open central vowel
5th	Haməs	[e]	Mid-high front vowel
6th	Sadəs	[ɨ] or Ø	High-central unrounded / epenthetic vowel
7th	Sabə	[o]	Mid-high back rounded vowel

The sixth order vowel [ɨ] is often epenthetic, inserted to break consonant clusters. A TTS model must decide whether a 6th order character is an audible [ɨ] or a bare consonant.

---

2. The Neural Synthesis Pipeline

flowchart TD
    A["Raw Ge'ez Text:<br/>'እንኳን ደህና መጡ'"] --> B["Orthographic Text Normalizer<br/>(Expand Numbers, Dates, Currency)"]
    B --> C["Contextual G2P & Gemination Tagger<br/>(Maps to IPA Phonemic Stream)"]
    C --> D["Neural Acoustic Model<br/>(VITS / FastSpeech 2 Duration & Pitch Predictors)"]
    D --> E["Latent Flow / Mel-Spectrogram"]
    E --> F["HiFi-GAN Neural Vocoder"]
    F --> G["Audio Waveform"]

VITS

Two-stage systems generate mel-spectrograms before a separate vocoder. VITS (Kim et al., 2021) is end-to-end:

• It integrates a text encoder, a stochastic duration predictor, and a variational autoencoder (VAE) with normalising flows.
• During training, it applies adversarial discriminator losses (derived from HiFi-GAN) directly to raw waveforms.