Spaces:
Running
Running
| """Derive ASR tasks from one FLEURS row. References never enter a public payload.""" | |
| from .schema import ( | |
| FAMILIES, | |
| character_scored, | |
| normalize_for_scoring, | |
| recording_id, | |
| task_id, | |
| ) | |
| SAMPLING_RATE = 16_000 | |
| # FLEURS utterances are read sentences; anything far outside this is a source defect | |
| # rather than a hard example, and is excluded rather than quietly scored. | |
| MIN_SECONDS = 0.5 | |
| MAX_SECONDS = 60.0 | |
| PROMPTS = { | |
| "transcription": ( | |
| "Transcribe this {name} speech. Return only the transcription, in {name}, " | |
| "lowercased and without punctuation." | |
| ), | |
| "verbatim_transcription": ( | |
| "Transcribe this {name} speech exactly as spoken, in {name}, preserving " | |
| "capitalization, punctuation, and numerals. Return only the transcription." | |
| ), | |
| "language_id": ( | |
| "Identify the language of this speech. Return only its FLEURS code, for " | |
| "example en_us or hi_in." | |
| ), | |
| } | |
| def derive_tasks(row, language, revision, *, families=FAMILIES, metadata_only=False): | |
| """Return every eligible task for one utterance, plus why any were skipped.""" | |
| skipped = {} | |
| def skip(reason): | |
| skipped[reason] = skipped.get(reason, 0) + 1 | |
| return [], skipped | |
| sample_id = row.get("id") | |
| if sample_id is None: | |
| return skip("missing_id") | |
| try: | |
| # FLEURS' id is a sentence id shared by several speakers; the recording is what | |
| # identifies one clip. | |
| recording = recording_id(row.get("path")) | |
| except ValueError: | |
| return skip("missing_path") | |
| num_samples = row.get("num_samples") or 0 | |
| duration = num_samples / SAMPLING_RATE | |
| if not MIN_SECONDS <= duration <= MAX_SECONDS: | |
| return skip("duration_outside_policy") | |
| audio = row.get("audio") or {} | |
| media = audio.get("bytes") if not metadata_only else None | |
| if not metadata_only and not isinstance(media, bytes): | |
| return skip("missing_audio_bytes") | |
| references = { | |
| "transcription": row.get("transcription"), | |
| "verbatim_transcription": row.get("raw_transcription"), | |
| "language_id": language, | |
| } | |
| tasks = [] | |
| for family in families: | |
| reference = references.get(family) | |
| if not isinstance(reference, str) or not reference.strip(): | |
| skipped[f"empty_{family}_reference"] = ( | |
| skipped.get(f"empty_{family}_reference", 0) + 1 | |
| ) | |
| continue | |
| if family != "language_id" and not normalize_for_scoring(reference, family): | |
| # Nothing would remain to score against. | |
| skipped[f"unscorable_{family}_reference"] = ( | |
| skipped.get(f"unscorable_{family}_reference", 0) + 1 | |
| ) | |
| continue | |
| tasks.append( | |
| { | |
| "task_id": task_id(revision, language, row["split"], recording, family), | |
| "language": language, | |
| "language_name": row.get("language") or language, | |
| "split": row["split"], | |
| "family": family, | |
| "sample_id": int(sample_id), | |
| "recording": recording, | |
| "prompt": PROMPTS[family].format(name=row.get("language") or language), | |
| "reference": reference, | |
| "media": media, | |
| "mime": "audio/wav", | |
| "sampling_rate": SAMPLING_RATE, | |
| "num_samples": int(num_samples), | |
| "duration_seconds": round(duration, 3), | |
| "error_unit": "cer" if character_scored(language) else "wer", | |
| "gender": row.get("gender"), | |
| } | |
| ) | |
| return tasks, skipped | |