deaditem/data-voice-vietnamese-restaurant-quan-oc

Vietnamese Restaurant Order Speech

This dataset contains Vietnamese spoken restaurant orders paired with text transcripts. Each utterance typically includes a table number, item quantities, dishes, drinks, and add-ons.

Dataset Structure

Files are split into subdirectories by filename-derived speaker_code to satisfy Hugging Face repository file-count limits:

• —metadata.csv: one row per audio sample.
• —audio/{speaker_code}/*.wav: mono WAV audio files.
• —transcripts/{speaker_code}/*.txt: original transcript files.

Columns in metadata.csv:

• —file_name: relative path to the WAV file, required by the Hugging Face folder-based audio builder.
• —transcription: Vietnamese transcript for the utterance.
• —transcript_file: relative path to the original TXT transcript.
• —source_id: source identifier parsed from the filename.
• —variant: filename variant, one of var0, var1, var2.
• —speaker_code: filename suffix such as NM1, NM2, NF, SM, or SF.
• —duration_seconds: audio duration in seconds.
• —sample_rate: WAV sample rate.
• —channels: number of audio channels.

Statistics

• —Audio/transcript pairs: 7,767
• —WAV files: 7,767
• —TXT files: 7,767
• —Total duration: approximately 15.64 hours
• —Average clip duration: approximately 7.25 seconds
• —Audio format: 24 kHz mono PCM WAV

Variant distribution:

• —var0: 2,589 samples
• —var1: 2,589 samples
• —var2: 2,589 samples

Speaker-code distribution across audio samples:

• —SM: 1,591 samples
• —NF: 1,588 samples
• —NM2: 1,550 samples
• —SF: 1,549 samples
• —NM1: 1,489 samples

The exact meaning of the speaker-code suffixes is not documented in the source archive, so treat them as filename-derived labels unless separately verified.

Loading

from datasets import Audio, load_dataset

ds = load_dataset("EmilyNguyen235/data-voice-vietnamese-restaurant")
# The folder-based builder exposes the WAV path as the `audio` feature.
ds = ds.cast_column("audio", Audio(sampling_rate=24000))

License and Data Notes

The source archive did not include a license file or dataset documentation. Verify usage rights before redistributing, training, or publishing derived models.