How TTS Pipelines Handle Geminate Consonants

Modern Text-to-Speech (TTS) systems must accurately synthesize geminate (doubled or prolonged) consonants to ensure intelligibility and naturalness in quantity-sensitive languages such as Italian and Finnish. This article examines how modern TTS pipelines process orthographic geminates, tracing their journey through text normalization, grapheme-to-phoneme (G2P) conversion, acoustic duration modeling, and neural vocoding to produce phonemically distinct long consonants.

Phonemic Significance of Geminates

In both Italian and Finnish, consonant length is phonemically contrastive; altering duration changes the meaning of a word.

• In Italian, the distinction separates words like fato (fate) and fatto (fact), or pala (shovel) and palla (ball). Italian also exhibits raddoppiamento fonosintattico, where word-initial consonants geminate after specific lexical and stress environments across word boundaries.
• In Finnish, consonant quantity is structurally embedded throughout the grammar, distinguishing minimal pairs like tuli (fire), tulli (customs), and tuuli (wind), with gemination occurring across nearly all consonant classes.

Because gemination directly impacts lexical identity rather than just stylistic cadence, TTS engines cannot treat double letters as simple orthographic redundancies.

Front-End Processing: Grapheme-to-Phoneme (G2P)

The TTS front-end resolves raw text into phonetic transcriptions. Handling orthographic geminates relies on how the G2P system interprets double consonants:

1. Phonetic Representation: Most TTS front-ends represent geminates either by using the International Phonetic Alphabet (IPA) length mark (e.g., /tː/) or by emitting two identical phonetic tokens sequentially (e.g., /t/ /t/). The explicit length mark /ː/ is generally preferred in modern systems because it treats the sound as a single prolonged segment rather than two separate articulations.
2. Rule-Based and Neural G2P: For Finnish, where orthography is highly transparent, rule-based mappings or finite-state transducers (FSTs) reliably map double letters directly to long phonemes. Italian requires hybrid approaches: while internal orthographic geminates (anno → /ˈan.no/ or /ˈanːo/) are straightforward, raddoppiamento fonosintattico (e.g., a casa pronounced as /a kˈka.sa/) requires context-sensitive neural G2P models or syntactic parsing rules to insert gemination markers at word boundaries.

Acoustic and Duration Modeling

Once phonemes are determined, the acoustic model converts the sequence into time-aligned representations, typically mel-spectrograms. The handling of geminates at this stage depends on the architecture:

• Explicit Duration Predictors (Non-Autoregressive Models): In architectures like FastSpeech 2, VITS, or Glow-TTS, an explicit duration predictor assigns a frame length to each input token. When trained on aligned audio corpora of Italian or Finnish, the duration predictor learns that phonemes tagged with a length marker require substantially higher frame counts (often 1.5 to 2.5 times the length of their singleton counterparts).
• Internal Alignment (Autoregressive Models): In older sequence-to-sequence models like Tacotron 2, attention mechanisms dynamically align phonemes to spectrogram frames. For geminates, the attention mechanism learns to dwell on the consonant's token across multiple decoding steps, generating an extended period of silence (for stops) or prolonged resonance (for fricatives, nasals, and liquids).

For stop consonants (such as /p/, /t/, /k/), the acoustic model primarily extends the silent closure phase prior to the release burst, rather than extending the burst itself. For continuous consonants (such as /s/, /m/, /l/), the model sustains the steady-state formant structure or frication noise over the extended duration.

Neural Vocoding and Waveform Generation

The final stage uses a neural vocoder (such as HiFi-GAN or diffusion-based vocoders) to convert the generated mel-spectrogram into raw audio.

Because the acoustic model provides the vocoder with a spectrogram containing an extended closure or extended continuous energy, the vocoder reconstructs the precise temporal phase of the sound. It preserves the clean, silent hold required for geminate stops without introducing phase artifacts or unnatural breathiness, followed by a crisp burst that signals the end of the lengthened constriction to the human ear.