Asaxi Prosody Recording Corpus

Navigation:

• The Asaxi Language Index
• Prosody, Stress and Intonation
• Clean reader text
• Interlinear reader text

The generated recording corpus turns the current Asaxi dictionary, grammar examples, dated elicitation controls, and reader text into a reproducible set for measuring timing and prosody. It is stored at:

99_Tools/festvox/corpora/asaxi-prosody-v1

The human-readable entry point is reader_corpus.md. Each item contains:

1. plain Asaxi text to read aloud;
2. an English translation, or a clearly labeled broader source context where the source paragraph cannot be split safely;
3. one pitch-accent row per written word;
4. the dictionary or fixed-expression H/L pattern;
5. the current model’s utterance-level H/L prediction;
6. separately labeled attested or dictionary reference evidence;
7. an explicit agreement result, source provenance, and any review warning.

H and L label relative mora-level pitch targets. They are not exact frequencies. A lexical pattern and its sentence realization can differ because of phrase accent, downstep, focus, deaccenting, or a boundary tone.

Annotation authority

The manifest keeps prediction and evidence separate. pitch_analysis.predicted always means output from the current rules. pitch_analysis.reference.authority: attested means the reading was transcribed from the dated elicitation appendix. dictionary means the reference is lexical or fixed-expression data rather than an utterance recording. The agreement record states which layer was compared. model_hypothesis means the current rules generated the reading and it still requires recording or linguistic review.

Unknown inflected forms and ambiguous homographs remain marked in the corpus. They are not silently accepted as ground truth.

The initial model matches all 11 dated controls whose reference has the same number of H/L values as the written form. The shorthand references for ŕoŕo daohè! and haśùnáhè! each omit one mora-level value, so they remain preserved and visibly marked mora_count_mismatch.

Recording and alignment

Record the Asaxi line only; the English translation is context and should not be spoken. Keep microphone position, speaking voice, room, sample rate, and gain stable. Use the requested recording ID as the WAV basename.

Keep the unedited recording. Store corrected word, mora, and phone boundaries, measured F0, and review decisions as separate annotation layers linked by the stable recording ID. This makes it possible to compare later models against the same speech rather than replacing the evidence.

The corpus is material for fitting and validating a prosody model. Its existence does not by itself verify acoustic timing, pitch realization, or naturalness.