Doubao-powered voice input for Linux, Wayland and Omarchy
"""Local speech samples and cancellable comparison, shared by the GTK UI."""
import json
from pathlib import Path
import threading
import uuid
import wave
from doubao_input.settings import write_atomic
from doubao_input.speech_eval import read_manifest, read_rows, read_audio, recognize, score
class SpeechWorkbench:
def __init__(self, directory):
self.directory = Path(directory)
self.samples = []
self.cancelled = threading.Event()
self.completed_results = []
path = self.directory / 'samples.jsonl'
if path.exists():
self.samples = read_rows(path)
if len(self.samples) > 200 or any(not isinstance(s.get("id"), str) or str(uuid.UUID(s["id"])) != s["id"] for s in self.samples):
raise ValueError("Invalid samples")
if len({s["id"] for s in self.samples}) != len(self.samples) or any(not isinstance(s.get("reference"), str) or any(not isinstance(s.get(k, []), list) or any(not isinstance(t, str) for t in s.get(k, [])) for k in ("terms", "protected_terms")) for s in self.samples):
raise ValueError("Invalid drafts")
def save(self):
self.directory.mkdir(parents=True, exist_ok=True, mode=0o700)
write_atomic(self.directory / 'samples.jsonl', ('\n'.join(json.dumps(s, ensure_ascii=False) for s in self.samples) + '\n').encode())
def add(self):
if len(self.samples) >= 200:
raise ValueError('Sample limit reached')
row = {'id': str(uuid.uuid4()), 'reference': '', 'terms': [], 'protected_terms': []}
self.samples.append(row)
self.save()
return row
def record(self, row, pcm):
if not pcm or len(pcm) > 1_920_000 or len(pcm) % 2:
raise ValueError('Invalid audio')
path = self.directory / (str(uuid.UUID(row['id'])) + '.wav')
self.directory.mkdir(parents=True, exist_ok=True, mode=0o700)
# Build WAV in memory so no recording is ever created world-readable.
from io import BytesIO
stream = BytesIO()
with wave.open(stream, 'wb') as audio:
audio.setnchannels(1); audio.setsampwidth(2); audio.setframerate(16000)
audio.writeframes(pcm)
write_atomic(path, stream.getvalue())
row['audio'] = str(path)
row.pop('results', None)
self.save()
def delete(self, row):
# Only remove audio owned by this sample, never an imported arbitrary path.
(self.directory / (str(uuid.UUID(row['id'])) + '.wav')).unlink(missing_ok=True)
self.samples.remove(row)
self.save()
def validated(self):
# UI permits incomplete drafts; validate the full set only when running.
from tempfile import TemporaryDirectory
with TemporaryDirectory() as temp:
path = Path(temp) / 'manifest.jsonl'
write_atomic(path, ('\n'.join(json.dumps(s, ensure_ascii=False) for s in self.samples) + '\n').encode())
rows = read_manifest(path)
for row in rows:
read_audio(row.get('audio', ''))
return rows
def compare(self, credentials, words, *, polisher=None, progress=lambda *_: None, recognizer=recognize):
rows = self.validated()
results = []
self.completed_results = results
for row in rows:
if self.cancelled.is_set():
break
pcm = read_audio(row['audio'])
for mode, hints in (('raw', []), ('hotwords', words)):
if self.cancelled.is_set():
break
progress(row['id'], mode)
if self.cancelled.is_set():
break
result = {'id': row['id'], 'mode': mode, **recognizer(pcm, credentials, hints, cancel=self.cancelled)}
if self.cancelled.is_set():
break
results.append(result)
if mode == 'hotwords' and polisher:
if self.cancelled.is_set():
break
import time
polished = {**result, 'mode': 'hotwords+polish'}
if not result.get('error'):
started = time.monotonic()
try:
polished['text'] = polisher(result['text'])
except (OSError, ValueError, TimeoutError):
polished['error'] = 'polish_failed'
polished['latency_ms'] += round((time.monotonic() - started) * 1000)
if not self.cancelled.is_set():
results.append(polished)
return results
def report(self, results):
return score(self.samples, results)