rinaldow/piper-onnx-durations

Piper ONNX Timing Models

Modified Piper TTS models with exposed duration tensors for temporal alignment.

Model Outputs

Output	Type	Description
audio	Float32Array	Synthesized speech
durations	Float32Array	Frame counts per token

Duration Format

Durations are expressed in audio frames. To convert to milliseconds:

duration_ms = duration_frames × (hop_size / sample_rate) × 1000

Standard values:

• hop_size: 256
• sample_rate: 22050 (medium models)

Token Structure

The durations array corresponds to a token sequence that includes structural markers:

ID	Type	Description
0	padding	Inter-token transitions
1	BOS	Beginning of sentence
2	EOS	End of sentence
3	space	Word boundary
4+	phoneme	Speech sounds

The token sequence is typically longer than the input phoneme string due to inserted structural tokens.

Configuration

Each model includes a .onnx.json file containing:

• phoneme_id_map: Symbol-to-ID mappings
• audio.sample_rate: Sample rate for duration conversion

Source

Based on rhasspy/piper-voices.

Important Notes on Model Selection

• Every model that was based on Lessac was excluded because they're not commercially usable.
• Every MLS (Multilingual LibriSpeech) model is skipped because they failed to produce short audio reliably in any language (just gibberish comes out if the sentence isn't long enough).