mirror of
https://github.com/got-feedBack/feedBack.git
synced 2026-07-28 15:42:35 +00:00
194 lines
7.8 KiB
Python
194 lines
7.8 KiB
Python
"""Per-syllable vocal pitch extraction via the demucs server's /pitch endpoint.
|
|
|
|
Sibling to `lyrics_transcribe.py` on the karaoke side: once we have
|
|
isolated vocals + per-syllable lyric timing (both produced by the
|
|
WhisperX fallback or shipped in the source archive), the /pitch endpoint
|
|
runs CREPE over the vocals stem and returns one MIDI note per supplied
|
|
timing token. The result lands in `<sloppak>/vocal_pitch.json` in the
|
|
shape the got-feedback/feedback-plugin-lyrics-karaoke renderer
|
|
already consumes:
|
|
|
|
{"version": 1, "notes": [{"t": float, "d": float, "midi": int}, ...]}
|
|
|
|
Pre-generating during sloppak conversion means the karaoke plugin no
|
|
longer has to run pYIN locally on every first-play — the file is
|
|
already there.
|
|
|
|
Engine selection
|
|
────────────────
|
|
This module exposes only the remote path (`extract_pitch_remote`). The
|
|
demucs server's /pitch endpoint runs CREPE on a GPU when available,
|
|
which is materially better than the pYIN fallback the karaoke plugin
|
|
runs locally. Adding a local CREPE path here would mean pulling
|
|
`crepe` + `tensorflow` as plugin deps (~500 MB+ on top of the
|
|
existing torch/demucs/whisperx). Deferred until users hit the gap.
|
|
If you need a local fallback today, install
|
|
`got-feedback/feedback-plugin-lyrics-karaoke` and let its local
|
|
pYIN run when the server isn't reachable.
|
|
|
|
Cache key parity with stem_separation / lyric_transcription
|
|
───────────────────────────────────────────────────────────
|
|
A `pitch_extraction` manifest block mirrors the shape introduced by
|
|
slopsmith#357: `{engine, model, version}`. Today engine is fixed at
|
|
`"crepe"` (the server's choice) and model at `"v1"` (server doesn't
|
|
yet expose the CREPE capacity dial it uses internally; this is the
|
|
requested value, same caveat as `lyric_transcription.model`). The
|
|
schema version is independent of the upstream CREPE version and bumps
|
|
per slopsmith's contract:
|
|
* patch — metadata-only or implementation fixes
|
|
* minor — backward-compatible additions
|
|
* major — output shape / semantics changed; existing
|
|
vocal_pitch.json should be regenerated and remote caches
|
|
should miss
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import logging
|
|
import math
|
|
from pathlib import Path
|
|
from typing import Callable, Optional
|
|
|
|
log = logging.getLogger("slopsmith.lib.vocal_pitch")
|
|
|
|
ProgressCB = Optional[Callable[[float, str, str], None]]
|
|
|
|
# `pitch_extraction` manifest-block constants. See module docstring for
|
|
# semver semantics.
|
|
PITCH_EXTRACTION_ENGINE = "crepe"
|
|
PITCH_EXTRACTION_MODEL = "v1"
|
|
PITCH_EXTRACTION_SCHEMA_VERSION = "1.0.0"
|
|
|
|
|
|
def extract_pitch_remote(
|
|
vocals_path: Path,
|
|
lyrics: list[dict],
|
|
server_url: str,
|
|
*,
|
|
api_key: str | None = None,
|
|
timeout: int = 300,
|
|
progress_cb: ProgressCB = None,
|
|
) -> list[dict]:
|
|
"""POST the vocal stem + lyric timings to `{server_url}/pitch`.
|
|
|
|
`lyrics` is the same `[{t, d, w}, ...]` list slopsmith writes to
|
|
`lyrics.json`. The endpoint only consumes `t` + `d` (it doesn't
|
|
need the word text), but we pass the full payload through —
|
|
slimmer to forward what we already have than to project.
|
|
|
|
Returns a normalized `notes` list: each entry is
|
|
`{"t": float, "d": float, "midi": int}` with `t`/`d` rounded to 3
|
|
decimals and `midi` coerced to int. Malformed server entries
|
|
(missing key, non-numeric, wrong type) are skipped rather than
|
|
propagated. Tokens the server couldn't extract a pitch for
|
|
(silent, sub-threshold confidence, no neighbour to borrow from)
|
|
are omitted from the response — the output may be shorter than
|
|
the input lyrics.
|
|
|
|
Errors raise `RuntimeError` with a truncated server response so
|
|
the caller can log+continue without bringing down the surrounding
|
|
transcription / split job. Transport-level failures
|
|
(`requests.RequestException`: connection, DNS, timeout, upload
|
|
aborted) are wrapped here so callers see one exception type, not
|
|
requests' hierarchy. Matches the failure idiom in
|
|
`transcribe_vocals_remote` and `_run_demucs_remote`."""
|
|
import requests
|
|
|
|
server_url = server_url.rstrip("/")
|
|
if progress_cb:
|
|
try:
|
|
progress_cb(0.10, "pitch", f"Uploading to CREPE server ({server_url})")
|
|
except Exception:
|
|
pass
|
|
|
|
headers: dict[str, str] = {}
|
|
if api_key:
|
|
headers["Authorization"] = f"Bearer {api_key}"
|
|
|
|
# The /pitch endpoint validates each entry has numeric `t` + `d`;
|
|
# the `w` field is ignored server-side but harmless to forward.
|
|
# Wrap the upload so a connection / DNS / timeout failure becomes a
|
|
# RuntimeError instead of leaking requests' own exception hierarchy
|
|
# past the docstring contract.
|
|
log.debug("POST %s/pitch vocals=%s lyrics=%d timeout=%ds",
|
|
server_url, vocals_path.name, len(lyrics), timeout)
|
|
# Catch both `requests.RequestException` (network / DNS / timeout /
|
|
# upload aborted) AND `OSError` from the `open(vocals_path)` itself
|
|
# (file disappeared between gate-check and upload, permissions
|
|
# change, EIO). Both surface as `RuntimeError` to keep the
|
|
# docstring contract one-type and let the caller handle every
|
|
# failure mode with one `except`. RequestException MUST be caught
|
|
# first because it inherits from OSError — listing OSError first
|
|
# would steal the network-failure path and label it as a vocals
|
|
# read error.
|
|
try:
|
|
with open(vocals_path, "rb") as f:
|
|
resp = requests.post(
|
|
f"{server_url}/pitch",
|
|
files={"file": (vocals_path.name, f, "audio/ogg")},
|
|
data={"lyrics": json.dumps(lyrics)},
|
|
headers=headers or None,
|
|
timeout=timeout,
|
|
)
|
|
except requests.RequestException as e:
|
|
raise RuntimeError(f"CREPE server request failed: {e}") from e
|
|
except OSError as e:
|
|
raise RuntimeError(f"Reading vocals stem {vocals_path.name} failed: {e}") from e
|
|
|
|
if resp.status_code != 200:
|
|
raise RuntimeError(
|
|
f"CREPE server error ({resp.status_code}): {resp.text[:300]}"
|
|
)
|
|
|
|
try:
|
|
data = resp.json()
|
|
except ValueError as e:
|
|
raise RuntimeError(f"CREPE server returned non-JSON: {e}") from e
|
|
|
|
if not isinstance(data, dict) or "notes" not in data:
|
|
raise RuntimeError(
|
|
f"CREPE server returned unexpected shape: {str(data)[:300]}"
|
|
)
|
|
|
|
raw_notes = data.get("notes")
|
|
if not isinstance(raw_notes, list):
|
|
raise RuntimeError(
|
|
f"CREPE server `notes` is not a list: {type(raw_notes).__name__}"
|
|
)
|
|
|
|
# Defensive: skip malformed entries entirely rather than crashing
|
|
# the whole pass on one bad record. Same posture as the WhisperX
|
|
# remote path. Non-finite t/d (NaN, ±Inf — a misbehaving server or
|
|
# a numerical edge in CREPE could surface them) are also filtered
|
|
# so they can't reach the on-disk vocal_pitch.json and break
|
|
# strict-JSON consumers downstream.
|
|
out: list[dict] = []
|
|
for n in raw_notes:
|
|
if not isinstance(n, dict):
|
|
continue
|
|
if "t" not in n or "d" not in n or "midi" not in n:
|
|
continue
|
|
try:
|
|
t = float(n["t"])
|
|
d = float(n["d"])
|
|
if not math.isfinite(t) or not math.isfinite(d):
|
|
continue
|
|
out.append({
|
|
"t": round(t, 3),
|
|
"d": round(d, 3),
|
|
"midi": int(n["midi"]),
|
|
})
|
|
except (TypeError, ValueError):
|
|
continue
|
|
|
|
log.debug("CREPE /pitch returned %d raw notes, %d after normalization",
|
|
len(raw_notes), len(out))
|
|
if progress_cb:
|
|
try:
|
|
progress_cb(1.0, "pitch", f"Got {len(out)} pitch notes")
|
|
except Exception:
|
|
pass
|
|
|
|
return out
|