Encoding

Ordinary encoding uses the configured voice map and preserves Piper’s framing:

BOS, PAD, (phoneme IDs, PAD)*, EOS

A map value may contain multiple IDs. The loaded voice’s phoneme_id_map is authoritative; there is no universal symbol vocabulary. PhonemizeResult.token_ids is a flattened convenience view. For Piper inference, use each sentence’s IDs separately.

Missing symbols

A produced phoneme absent from the configured map is recorded in EncodeResult and sentence/result data. The missing policy controls whether processing stops, warns, or continues silently:

from piperg2p import MissingPhonemePolicy, phonemize_prepared

result = phonemize_prepared(
    "Hello world",
    language="en-us",
    config="voice.onnx.json",
    missing=MissingPhonemePolicy.WARN,
)
print(result.missing_phonemes)

Policy	Behavior
`MissingPhonemePolicy.ERROR`	Raise `MissingPhonemeError`
`MissingPhonemePolicy.WARN`	Emit `MissingPhonemeWarning` and skip the unmapped symbol
`MissingPhonemePolicy.IGNORE`	Skip the unmapped symbol silently

With WARN and IGNORE, the absent symbol remains recorded in result data even though encoding continues.