Pavan-salient/updated-marshall-asr

Westlake Marshall Customer ASR Benchmark

This dataset packages 98 customer-only Westlake Marshall calls that were manually annotated by human reviewers. It is designed for benchmarking automatic speech recognition (ASR) systems with word error rate (WER) metrics on real customer service traffic.

Source Data

• Audio: Dual-channel MP3 recordings retrieved from the Westlake Marshall servicing platform. The customer channel (left) was isolated and converted to mono 16 kHz WAV for each customer speaking turn.
• Transcripts: Human-written transcriptions focused on the customer turns. Transcripts may include * markers where words were unintelligible.

Two calls present in the original annotation sheet were missing audio assets and are excluded from this release.

Dataset Structure

westlake_asr/
  audio/
    <call_id>/
      turn_<index>.wav
  data/
    turns.jsonl
    stats.json
  README.md
  (prep scripts live under `src/stt_eval/` in the repo)

Turn Record Schema (turns.jsonl)

Each line is a minimal JSON object with:

Field	Type	Description
call_id	string	Call UUID
turn_index	int	1-based index within the call
text	string	Human transcript for the turn (raw casing/punctuation)
normalized_text	string	Lowercased alphanumeric text used as the WER reference
needs_manual_review	bool	true when the transcript contained * (flag for follow-up)
audio_filepath	string	Relative path to the cropped 16 kHz mono WAV clip
duration_ms	int	Length of the exported clip in milliseconds

Audio clips live under audio/; Hugging Face renders them via the dataset viewer directly.

stats.json summarizes aggregate totals (530 turns; 53 flagged for manual review, approx. 10%).

Quality Considerations

• Timestamps originated from human annotations and may drift by ±1 second. Each clip has a 100 ms leading pad and 200 ms trailing pad to compensate.
• Clips shorter than 400 ms are automatically extended (when possible) to ensure compatibility with most ASR models.
• Use the needs_manual_review flag to filter or separately analyze segments containing unintelligible speech.

Suggested Evaluation Workflow

1. Push any edits to Hugging Face (scripts/push_hf_dataset.py …).
2. Run the evaluation command, which always pulls the dataset from Hugging Face (default repo TrySalient/marshall-asr) to guarantee you are measuring against the published 12-second audio version.
3. For each record, the script feeds the WAV clip to the ASR engine and compares its output to clean_text using WER (jiwer).
4. Reports include overall WER plus clean vs. flagged breakdowns, and the failure artifacts are written under datasets/westlake_asr/data/evaluations/ for review.

Review / Re-annotation Workflow

• Launch the local review tool:

  source .venv_tts_eval/bin/activate
  streamlit run tools/review_app.py
  

  Use the interface to listen to clips, update transcripts, and toggle the needs_manual_review flag. Saved changes appear in the “Pending edits” table and can be exported to outputs/review_exports/review_edits_TIMESTAMP.json.

• Apply edits to turns.jsonl:

  python scripts/apply_review_edits.py --edits outputs/review_exports/review_edits_TIMESTAMP.json
  

• After applying, regenerate evaluation metrics if needed and push the dataset back to Hugging Face with scripts/push_hf_dataset.py.

Licensing and Usage

This dataset contains proprietary Westlake Marshall customer content and is provided for internal benchmarking only. Do not redistribute outside authorized Salient/Skwid environments.