Args:
    seconds: recording length (max 120 by default). Tell the user before calling.
    name: optional file name (without extension); defaults to date and time.

Returns `audio` (path of the .wav), levels (peak/RMS in dBFS, estimated SNR, % clipping),
`confidence` and `warnings`. The `no_microphone_permission` warning means macOS delivered no
audio: the user must enable the microphone for the app in System Settings.

Args:
    audio: audio path (wav/flac/mp3/m4a) or a name returned by record_audio.
    reference_midi: optional path to a .mid with the expected melody. Tolerates singing an
        octave up/down, starting at another moment and a somewhat different tempo.
    key: optional when there is no MIDI, e.g. "A minor", "G major", "F#" (Spanish names such as
        "La menor" also work): each note is compared with the nearest note of that scale.
        Without it, with the nearest equal-tempered note.
    midi_track: index of the MIDI track to use (default: the non-drum track with most notes).

Returns `summary` (notes, range, pct_within_25_cents, signed mean_deviation_cents,
tendency flat/sharp/centered, drift_cents between 1st and 2nd half, stability counts; with
MIDI also missed, extra_notes, wrong_notes, onsets on_time/late/early, octave_shift,
start_offset_ms and relative_tempo >1 = slower) and `notes`: t (s), dur, note or
expected/sung, cents (+ = sharp), stability (stable | vibrato with vibrato_hz and
vibrato_pp_cents | unstable with std_cents), drift_cents within the note and, with MIDI,
onset_ms/onset; reference notes not sung have status=missed and unmatched sung notes
status=extra. If the list does not fit, the most problematic notes are shown and the rest
is in `full_details`.

Args:
    audio: audio path or a name returned by record_audio.
    language: "en", "es", "en-US", "Spanish"…; auto-detected when omitted.

Returns `language`, `language_prob`, `text`, `segments` [{t, end, text}] and `words`
[{w, t, end}]. If it does not fit, words are dropped first and then segments (everything
stays in `full_details`). The first call downloads the model (VOICE_COACH_WHISPER_MODEL).

Args:
    audio: audio path or a name returned by record_audio.
    language: language of the speech ("en" or "es"; filler lists exist for both).
    pause_threshold_s: minimum silence between words to count as a long pause.

Returns:
- `pace`: words, speaking_time_s, words_per_minute, wpm_excluding_long_pauses.
- `pauses`: count, total_s, max_s, per_minute and items [{t, dur, after: previous word}].
- `fillers`: total, per_minute, by_type and occurrences [{t, text}]. Ambiguous words
  ("like", "so") only count when isolated by a comma or a pause.
- `pitch`: median_f0_hz, std_semitones, range_p10_p90_semitones, monotone (std < 2 st) and
  monotone_stretches [{t, end}] over 8 s windows.
- `volume`: mean_dbfs, std_between_words_db, range_p10_p90_db, low_variation (std < 3 dB).
- `segments` of the transcript [{t, end, text}].

Args:
    audio: audio path or a name returned by record_audio (up to 60 s; short sentences work best).
    expected_text: the exact text that was read.
    language: "en-US" (default) or "en-GB".

Returns `summary` (overall_score 0–100, words_with_errors, error counts by type,
frequent_confusions expected→detected) and `words`: word, score, t (s) and, when there are
errors, expected/detected in IPA and errors [{type: substitution|deletion|insertion,
expected, detected}]; `missing` when the word was not said. `phonemes` holds both full
sequences. Word score = 100·(1 − (S + D + 0.5·I)/expected phonemes). Allophonic variants
(flap, schwa, linking r) do not count as errors. The text_mismatch warning means a
different text was probably read. Backend: local wav2vec2, or Azure when
VOICE_COACH_PRONUNCIATION_BACKEND=azure (then the audio is sent to Azure).

Use it at the start of a session to pick up the previous focus. Returns `file`,
`total_entries` and `entries` (markdown of each session, oldest to newest).

Args:
    entry: markdown of the session. It should start with "## YYYY-MM-DD · mode · exercise";
        if it does not start with "## ", the date and time are added as the title.