instinct-org/espeech_podcasts_chunked_nfa_aligned

Forced-Aligned STT Dataset

Source dataset: instinct-org/espeech_podcasts_chunked

Aligned dataset: instinct-org/espeech_podcasts_chunked_nfa_aligned

Rows: 231121 successfully aligned rows

This dataset adds NeMo Forced Aligner metadata for STT training and timestamp quality control. Source rows that did not produce usable CTM alignments are excluded from the published data.

Added columns:

• nfa_token_alignments: NeMo token/subword CTM spans
• nfa_word_alignments: word-level CTM spans
• nfa_segment_alignments: segment-level CTM spans
• nfa_character_alignments: character spans; true only when the model token CTM is character level, otherwise derived from word spans
• nfa_character_alignment_source
• nfa_true_character_alignment
• nfa_alignment_model
• nfa_alignment_status
• nfa_alignment_error

NeMo Forced Aligner emits token, word, and segment CTM levels for this model family. With BPE/subword ASR models, token level is not the same as character level.

Model: nvidia/stt_uz_fastconformer_hybrid_large_pc

Quarantine Notice

This forced-aligned dataset is quarantined and should not be used for STT training.

Reason:

• The source dataset is Russian (language: ru), but this alignment run used nvidia/stt_uz_fastconformer_hybrid_large_pc.
• Only the first 11 source Parquet shards (train-00000 through train-00010) contain embedded audio.bytes.
• Source shards from train-00011 onward contain audio.bytes = null and Windows-local file paths such as D:\ds\..., which are not accessible in the training environment.
• As a result, only 231,121 rows were aligned and 2,565,957 rows were excluded.

Next action: repair or regenerate the source dataset so all rows contain embedded audio bytes or accessible audio files, then rerun forced alignment with a Russian-capable or multilingual aligner. Do not include this repo in STT training data composition until it is regenerated.