How Gemination Shapes Amharic Speech & NLP

In Italian, the contrast between a singleton and a geminate consonant separates destiny (fato [ˈfaːto]) from a deed (fatto [ˈfatːo]). In Arabic, the diacritic shaddah (ـّ) marks a doubled consonant.

In Amharic, gemination also changes meaning, but standard Ge’ez orthography does not write it. Two different words can share one spelling.

1. Minimal Pairs

Gemination prolongs the closure or constriction phase of a consonant.

Lexical Item	Orthography	Phonemic Tier	Syntactic & Semantic Category
አለ (alä)	`አለ`	`/ʔalɐ/`	Verb of quotation (“He said”)
አለ (allä)	`አለ`	`/ʔalːɐ/`	Existential defective copula (“There is / He exists”)
ገና (gäna)	`ገና`	`/gɐnɐ/`	Temporal adverb (“Still / Not yet”)
ገና (Gänna)	`ገና`	`/gɐnːɐ/`	Proper noun (“Genna / Ethiopian Christmas”)

2. Morphological Gemination: Type A, Type B, and Frequentatives

Amharic triliteral verbs fall into conjugational types by their stem gemination patterns:

Type A Verbs (e.g., ሰበረ säbbärä): The second radical (C2) is geminated in the perfective aspect (säbbärä “he broke”), but becomes a singleton in the imperfective (yəsäbrall “he breaks”).
Type B Verbs (e.g., ፈለገ fällägä): The second radical is permanently geminated across all aspects and moods (fällägä “he wanted”, yəfälləgall “he wants”, məfälläg “to want”).

Frequentative Reduplication: Frequentative verb stems reduplicate the second radical with a 4th order vowel (-a-), followed by gemination:

Root:        ሰ-ብ-ር (s-b-r)
Imperfective: ይሰብራል (yəsäbrall)     -> Medial C2 is singleton
Frequentative: ይሰባብራል (yəsäbabbərall) -> Medial C2 reduplicated and geminated

Character bigrams cannot tell Type A from Type B. That takes lemma identification.

3. Text-to-Speech: Grapheme-to-Phoneme (G2P)

In VITS, FastSpeech 2 or Matcha-TTS, a duration predictor processes the text before the decoder and vocoder.

flowchart TD
    A["Raw Ge'ez Orthography:<br/>'ንጉሡ አለ' vs 'ንጉሡ ሰላም አለ'"] --> B["Contextual Morpho-Syntactic Disambiguator<br/>(Transformer Sequence Tagger)"]
    B --> C["Explicit Phonemic Tier with Gemination<br/>/nɨgusu ʔalːɐ/ vs /nɨgusu sɐlam ʔalɐ/"]
    C --> D["Neural Duration Predictor"]
    D --> E["Flow-Matching / Latent Diffusion<br/>(VITS / Matcha-TTS)"]
    E --> F["Neural Vocoder (HiFi-GAN)"]
    F --> G["Natural Waveform Audio"]

Trained on raw Ge’ez text without a G2P layer, the duration predictor sees አለ with both durations. Minimizing mean squared error pulls its prediction between them, and the output sounds like neither word.

The fix is an intermediate phonemic representation with explicit gemination markers (e.g., the length mark ː or doubled IPA symbols):

def disambiguate_ale(prev_tokens: list[str]) -> str:
    """
    Heuristic syntactic check for the minimal pair 'አለ':
    - If preceded by direct speech quotes or conversational markers -> /ʔalɐ/ (said)
    - If functioning as existential copula -> /ʔalːɐ/ (there is)
    """
    quote_markers = {"ብሎ", "በማለት", "ሰላም", "እንኳን", "እንዲህ"}
    if any(marker in prev_tokens[-3:] for marker in quote_markers):
        return "ʔalɐ"  # Singleton: 'said'
    return "ʔalːɐ"     # Geminate: 'there is'

4. Speech Recognition: Hearing the Contrast

An acoustic model must detect the durational difference to segment phrases correctly. Two variables degrade this:

Speaking Rate: A geminate in fast speech and an emphatic singleton in slow speech can overlap in duration.
Audio Compression: Telephony codecs (G.711, GSM-FR) and lossy MP3 compression blur spectral transitions, masking the silent closure phase of ejective and stop geminates (kː, tː, p'ː).

Speed perturbation and SpecAugment time-masking are the usual training augmentations for these conditions.