munation/amharic-tts-blindspot

Amharic same-spelling gemination: a blind spot in open TTS

What I noticed (Question 1)

As an Amharic speaker, I keep noticing words that are written exactly the same but mean different things depending on how people say them. In school we called this ጠብቆ vs ላልቶ. In linguistics this is gemination vs non-gemination: the consonant is held longer, or it is not. The writing does not show it.

Note on labels: in this README and in file names, “stressed” means gemination (a held / doubled consonant = ጠብቆ). “Relaxed” means non-gemination (ላልቶ). This is not English syllable stress (louder syllable).

A clear everyday example is አለ:

• አለ said ላልቶ (relaxed / non-gemination) → “he said”
• አለ said ጠብቆ (stressed / gemination) → “there is / exists”

Same letters. Different meaning. Only speech carries the difference.

I have heard TTS systems sound fluent in Amharic but still miss this. They read the letters and pick one default sound. That is the blind spot I tested: sounds that change the meaning of a word, while the writing stays the same.

So a TTS model cannot “see” the answer in the spelling. It has to use the sentence meaning to pick gemination vs non-gemination (ጠብቆ vs ላልቶ).

Benchmarks usually check if speech sounds clear. They rarely check this: same letters, two meanings — does the model switch the pronunciation?

File names use stressed / relaxed so folders stay easy to sort. They mean gemination / non-gemination, not extra letters on the page.

Model

I used Lab-et/omnivoice-amharic (~0.6B parameters).

Why this model:

• open Amharic TTS
• fits the challenge size range (0.6B–6B)
• runs in Google Colab

What I did

I built 4 pairs. Each pair is one spelling, two meanings (gemination / ጠብቆ vs non-gemination / ላልቶ). For each meaning I wrote a short context sentence myself.

Same spelling	Gemination (stressed / ጠብቆ)	Non-gemination (relaxed / ላልቶ)
ዋና	key / important	swim
ሰፊ	wide	tailor's shop
ገና	gena (Christmas game)	yet / still
አያት	grandparent	looked intently

Steps:

1. Generate TTS for all 8 sentences (first pass).
2. Listen as a native speaker and label each clip as stressed (gemination) or relaxed (non-gemination).
3. Regenerate each sentence 5 times (40 clips) to check if the failure is stable or random. Each of the 5 clips per sentence is a separate model generation (a different audio file), not a copy of the same file. I did this so the result does not depend on a single run.
4. Record my own human reference audio for all 8 sentences (so you can hear the contrast the model misses).

Full sentences and labels are in results/.

Results

First pass: in all 4 pairs, both context sentences got the same pronunciation (0/4 switched).

5× regeneration: every sentence stayed the same across all 5 separate generations (40/40 stable). Still, in all 4 pairs, both sentences kept that same pronunciation (0/4 switched).

What I heard:

• ዋና and ሰፊ → always gemination (stressed / ጠብቆ), in both contexts and all repeats
• ገና and አያት → always non-gemination (relaxed / ላልቶ), in both contexts and all repeats

So in this pilot, the model did not use sentence meaning to choose gemination vs non-gemination for these same-spelling pairs.

Limits (this is a pilot)

• Only 4 pairs
• One listener (me)
• Listening was not blind — I knew which sentence was which when I labeled
• I wrote the context sentences myself
• Labels from listening, not yet measured consonant duration
• One model

How to reproduce

• Google Colab
• Model: Lab-et/omnivoice-amharic via the OmniVoice package
• Language: Amharic
• Settings used: num_step=32, guidance_scale=2.0
• Generate each of the 8 sentences once, then regenerate each sentence 5 times
• Compare against the human recordings in audio/human_recordings/

Path forward (Question 3)

To fix this, we need training data that makes the choice explicit:

• same spelling (identical text)
• two meanings
• two gold audios (gemination vs non-gemination)
• clear context sentences for each

Then fine-tune the TTS so it must pick the right spoken form from context — not one default for both sentences of a pair.

A stronger option later: give the model an explicit control for this contrast (hold length / gemination).

What’s in this bucket

audio/human_recordings/       my spoken references (8)
audio/tts_single_generation/  first TTS outputs (8)
audio/tts_five_repeats/       5 regenerations per sentence (40)
results/dataset_pairs.txt     pairs, sentences, first-pass labels
results/listening_five_repeats.txt
results/summary.txt
docs/pair_names.md            file-name legend

Model: https://huggingface.co/Lab-et/omnivoice-amharic