Build 1 clip(s) from data/background-conversations/clips.json

README.md

@@ -0,0 +1,51 @@
+---
+license: cc-by-4.0
+language:
+- en
+task_categories:
+- automatic-speech-recognition
+pretty_name: Background Conversations
+tags:
+- speaker-isolation
+- target-speaker-asr
+- background-speech
+- test-set
+- hebrish
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: data/test-*.parquet
+---
+
+# Background Conversations
+
+<!-- dataset-card-frame:begin -->
+[![Type: Audio](https://img.shields.io/badge/Type-Audio-1f6feb)](https://huggingface.co/datasets/Hebrish/Dataset-Index) [![Role: Standalone](https://img.shields.io/badge/Role-Standalone-8250df)](https://huggingface.co/datasets/Hebrish/Dataset-Index)
+
+**Audio dataset, standalone.** Kept apart from the master dataset and its split-outs. Every audio dataset as one tree, with hours: [the audio index](https://huggingface.co/datasets/danielrosehill/Audio-Dataset-Index). How every dataset fits together: [the dataset index](https://huggingface.co/datasets/Hebrish/Dataset-Index).
+<!-- dataset-card-frame:end -->
+
+A small **test-only** set for target-speaker ASR: one person (the foreground speaker) dictates into a phone while other
+people talk, a baby cries and household devices make noise in the same room. The transcript is the foreground speaker's
+words only. Everything else is what a model, or a speaker-isolation front end before it, should leave out.
+
+It is never trained on. It is for checking how a dictation model, or a voice-isolation step, copes with real rooms.
+
+## Data
+
+- **1 clip(s), 3.1 min**, split `test` only.
+- Audio: FLAC, 48 kHz, 16-bit, mono, converted from phone voice notes (Opus).
+- `text`: foreground speech, verbatim, punctuated and cased. `transcript_status` is `pending` until the speaker has
+  checked it against the audio, then `reviewed`.
+- `model_hypothesis` / `asr_model`: the untouched first-pass transcript the draft started from.
+- `noise_flags`: what is audible behind the speaker (`background-speech`, `baby-crying`, `device-voice-prompt`, ...).
+- `extra` (JSON text): `background_sources` (who or what is in the background), `diarized` (false: the background speech
+  has no transcript or segment labels yet), `source_file`.
+
+Columns follow the Hebrish audio format standard (`audio.schema.json` in this repo). There is no `sentences` tab: these are
+free speech, not read sentences.
+
+## Licence
+
+CC-BY-4.0.

audio.schema.json

@@ -0,0 +1,599 @@
+{
+ "$schema": "https://json-schema.org/draft/2020-12/schema",
+ "$id": "https://huggingface.co/datasets/Hebrish/Example-Audio-Sentences/raw/main/audio.schema.json",
+ "title": "Hebrish audio dataset row (standard of 2026-11-06)",
+ "description": "Metadata of one take in a Hebrish audio dataset. The parquet also carries an `audio` column (struct<bytes: binary, path: string>), required in parquet and not listed here because metadata jsonl exports drop it. Source of truth: ASR-Training docs/dataset-format-standard.md.",
+ "type": "object",
+ "required": [
+  "id",
+  "text",
+  "duration_s",
+  "sample_rate",
+  "bit_depth",
+  "audio_format",
+  "categories",
+  "target_terms",
+  "prompt_id",
+  "prompt_source",
+  "mode",
+  "speaker",
+  "speaker_id",
+  "speaker_gender",
+  "contributor",
+  "recorded_at",
+  "device",
+  "microphone",
+  "environment",
+  "transcript_status",
+  "pronunciation",
+  "sentence_id"
+ ],
+ "properties": {
+  "id": {
+   "type": "string",
+   "description": "Take id <UTC timestamp>_<6 chars>, e.g. 20261104T221604Z_d9mx46."
+  },
+  "text": {
+   "type": "string",
+   "description": "Written form, the ASR target: digits, symbols. May be empty only while transcript_status is pending."
+  },
+  "duration_s": {
+   "type": "number",
+   "description": "Seconds of audio."
+  },
+  "sample_rate": {
+   "type": "integer",
+   "description": "48000 / 24 from 2026-10-22; legacy 16000 / 16."
+  },
+  "bit_depth": {
+   "type": "integer",
+   "description": "48000 / 24 from 2026-10-22; legacy 16000 / 16."
+  },
+  "audio_format": {
+   "enum": [
+    "flac",
+    "wav"
+   ],
+   "description": "flac from 2026-10-22; legacy wav."
+  },
+  "categories": {
+   "type": "array",
+   "items": {
+    "type": "string"
+   },
+   "description": "Category tags of the take, e.g. tech, code_switch, target_word."
+  },
+  "target_terms": {
+   "type": "array",
+   "items": {
+    "type": "string"
+   },
+   "description": "Terms the sentence exists to teach."
+  },
+  "prompt_id": {
+   "type": "string",
+   "description": "Id of the prompt in the bank (p_..., pw_..., a task prompt id)."
+  },
+  "prompt_source": {
+   "type": "string",
+   "description": "Where the prompt came from, e.g. on-demand, generated:<model>, queue:<site>."
+  },
+  "mode": {
+   "type": "string",
+   "description": "Recording mode in the app, e.g. bank, free."
+  },
+  "device": {
+   "type": "string"
+  },
+  "microphone": {
+   "type": "string"
+  },
+  "environment": {
+   "type": "string",
+   "description": "Room acoustics label, e.g. quiet-indoor."
+  },
+  "speaker": {
+   "type": "string",
+   "description": "Speaker's full name; training reads only Daniel Rosehill."
+  },
+  "speaker_gender": {
+   "type": "string"
+  },
+  "contributor": {
+   "type": "string"
+  },
+  "speaker_id": {
+   "type": "string",
+   "description": "Permanent random 4-digit speaker id (Hebrish/Speakers), as a string.",
+   "pattern": "^[0-9]{4}$"
+  },
+  "recorded_at": {
+   "type": "string",
+   "description": "UTC ISO 8601 timestamp."
+  },
+  "transcript_status": {
+   "enum": [
+    "script",
+    "pending",
+    "reviewed"
+   ],
+   "description": "script = read from the card; pending = free speech awaiting a transcript."
+  },
+  "pronunciation": {
+   "type": "array",
+   "description": "Reading guide: one object per term whose reading is unclear, derived from the Recorder's registry recorder/app/public/data/pronunciation.json (recorder/scripts/pronunciation.py). [] by default. Not a record of how a take was said (that is hebrew_pronunciation / pronunciation_variants).",
+   "items": {
+    "type": "object",
+    "required": [
+     "term",
+     "say",
+     "status"
+    ],
+    "additionalProperties": false,
+    "properties": {
+     "term": {
+      "type": "string",
+      "minLength": 1,
+      "description": "The term as it is spelled in text."
+     },
+     "say": {
+      "type": "string",
+      "pattern": "^[A-Za-z' ]+(-[A-Za-z' ]+)*$",
+      "description": "Stress respelling: syllables joined by hyphens, stressed syllable in CAPITALS, e.g. ah-ka-MOL. For an Israeli brand, place or institution, how it is said in Israel."
+     },
+     "note": {
+      "type": [
+       "string",
+       "null"
+      ],
+      "description": "Optional: why, or which reading is meant."
+     },
+     "status": {
+      "enum": [
+       "unreviewed",
+       "ai-reviewed",
+       "human-reviewed"
+      ]
+     },
+     "say_en": {
+      "type": [
+       "string",
+       "null"
+      ],
+      "pattern": "^[A-Za-z' ]+(-[A-Za-z' ]+)*$",
+      "description": "Optional: the usual reading in English-speaking countries when it differs from `say` (NYOO-ro-fen)."
+     },
+     "hebrew": {
+      "type": [
+       "string",
+       "null"
+      ],
+      "description": "Optional: the Hebrew-script spelling, the evidence for the reading (for drugs, as in the Ministry of Health registry). Never part of a transcript."
+     }
+    }
+   }
+  },
+  "text_he": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "text_full": {
+   "type": [
+    "string",
+    "null"
+   ],
+   "description": "Full-fidelity target when it differs from text; null otherwise."
+  },
+  "spoken_text": {
+   "type": [
+    "string",
+    "null"
+   ],
+   "description": "Exactly what was read."
+  },
+  "task": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "topic": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "style": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "style_instruction": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "prompt_origin": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "prompt_kind": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "prompt_note": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "prompt_set": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "term_language": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "paragraph_text": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "verbatim_text": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "aside_text": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "aside_position": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "noise_level": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "notes": {
+   "type": [
+    "string",
+    "null"
+   ],
+   "description": "Free note on the take. Blank in public reference rows."
+  },
+  "model_hypothesis": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "asr_model": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "hebrew_pronunciation": {
+   "type": [
+    "string",
+    "null"
+   ],
+   "description": "How a Hebrew term WAS said in this take (a record of the take, not a guide). Not the same as pronunciation."
+  },
+  "variant_spelling": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "imported_from": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "original_text": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "text_before_normalization": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "source_prompt_id": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "capture_mode": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "tts_split": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "session_id": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "room_tone_id": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "room": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "location": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "location_name": {
+   "type": [
+    "string",
+    "null"
+   ],
+   "description": "Place name for away takes. Blank unless the take is meant to carry it."
+  },
+  "location_note": {
+   "type": [
+    "string",
+    "null"
+   ],
+   "description": "Free note on the place. Blank in public reference rows."
+  },
+  "priority_word": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "extra": {
+   "type": [
+    "string",
+    "null"
+   ],
+   "description": "JSON text holding columns not in the typed schema (mic_name, activity, pronunciation_variants, ...)."
+  },
+  "source_url": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "ui_note": {
+   "type": [
+    "string",
+    "null"
+   ]
+  },
+  "conditions": {
+   "type": [
+    "array",
+    "null"
+   ],
+   "items": {
+    "type": "string"
+   }
+  },
+  "noise_flags": {
+   "type": [
+    "array",
+    "null"
+   ],
+   "items": {
+    "type": "string"
+   }
+  },
+  "variant_groups": {
+   "type": [
+    "array",
+    "null"
+   ],
+   "items": {
+    "type": "string"
+   }
+  },
+  "background_speech_languages": {
+   "type": [
+    "array",
+    "null"
+   ],
+   "items": {
+    "type": "string"
+   }
+  },
+  "prompt_categories": {
+   "type": [
+    "array",
+    "null"
+   ],
+   "items": {
+    "type": "string"
+   },
+   "description": "Categories of the source prompt. A list (live subsets store a string)."
+  },
+  "edited": {
+   "type": [
+    "boolean",
+    "null"
+   ]
+  },
+  "clipped": {
+   "type": "boolean"
+  },
+  "tts_eligible": {
+   "type": [
+    "boolean",
+    "null"
+   ]
+  },
+  "browser_processing": {
+   "type": "boolean"
+  },
+  "has_digits": {
+   "type": "boolean"
+  },
+  "fallback_ok": {
+   "type": [
+    "boolean",
+    "null"
+   ],
+   "description": "With text_full: whether text is still a correct transcript. A bool (live subsets store a string)."
+  },
+  "written_form_uncertain": {
+   "type": [
+    "boolean",
+    "null"
+   ]
+  },
+  "native_rate": {
+   "type": [
+    "integer",
+    "null"
+   ]
+  },
+  "clipped_samples": {
+   "type": [
+    "integer",
+    "null"
+   ]
+  },
+  "lead_trim_s": {
+   "type": [
+    "number",
+    "null"
+   ]
+  },
+  "peak_dbfs": {
+   "type": [
+    "number",
+    "null"
+   ]
+  },
+  "floor_dbfs": {
+   "type": [
+    "number",
+    "null"
+   ]
+  },
+  "speech_dbfs": {
+   "type": [
+    "number",
+    "null"
+   ]
+  },
+  "snr_db": {
+   "type": [
+    "number",
+    "null"
+   ]
+  },
+  "edge_head_dbfs": {
+   "type": [
+    "number",
+    "null"
+   ]
+  },
+  "edge_tail_dbfs": {
+   "type": [
+    "number",
+    "null"
+   ]
+  },
+  "location_accuracy_m": {
+   "type": [
+    "number",
+    "null"
+   ]
+  },
+  "latitude": {
+   "type": [
+    "number",
+    "null"
+   ],
+   "description": "Decimal degrees as a number, only for away places, null otherwise. Live subsets store a string."
+  },
+  "longitude": {
+   "type": [
+    "number",
+    "null"
+   ],
+   "description": "Decimal degrees as a number, only for away places, null otherwise. Live subsets store a string."
+  },
+  "pronunciation_variants": {
+   "type": [
+    "array",
+    "null"
+   ],
+   "description": "How a registry term WAS said in this take: [{term, variant, category}].",
+   "items": {
+    "type": "object",
+    "required": [
+     "term",
+     "variant"
+    ],
+    "properties": {
+     "term": {
+      "type": "string"
+     },
+     "variant": {
+      "type": "string"
+     },
+     "category": {
+      "type": [
+       "string",
+       "null"
+      ]
+     }
+    }
+   }
+  },
+  "sentence_id": {
+   "type": "string",
+   "pattern": "^p_[0-9a-f]{10}$",
+   "description": "The row's id in this dataset's sentences tab"
+  }
+ }
+}

data/test-20261112T162206Z-b90f.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b253df44d0d838d282bad65957fa3e13b33a6820f94dadd0c01ea264c9d08750
+size 5360427

meta/test.jsonl

@@ -0,0 +1 @@
+{"id": "20261112T161250Z_ce058f", "sentence_id": "p_02f106c45a", "text": "And recording a quick note here. And we're trying to do voice isolation. There is a man called Mike who's sitting on our couch. And this is a small, a very small dataset that it's very. It's an experimentary one for people speaking in the background. So there's all kinds of people. So I'm just going to describe who I am. My name is Daniel, I'm the speaker here recording on the OnePlus Nord 3 5G and other people in the room. There's a man called Mike who's looking very unwell, lying on the couch. An Australian man. Ezra is my son and he's just triggered the Roborock speaker. Hannah is my wife. And also made. It made a noise there. And we also had things. So this is a good first. First clip where we have. You can hear Ezra crying there. So this sample we have a challenge for background noise and for speaker isolation. We have non-diarized audio. So we're getting a few angles for this one here. And the purpose of this dataset, this is a very experimentary part of the project because one of the tasks that I wanted to train on for my ASR fine-tune was specifically speaker isolation. And as I've mentioned in many of these blog posts, the challenges that I encounter dictating all the time are basically in the real world. It's not a pristine studio. People are having conversations. There's like PA messages. That's what you have when you're dictating. And I thought firstly, naively that there could be just record clips, ground truth is your speech. And therefore the model will eventually learn to discard everything ranging from bar music to what have you. And as, as I've been checking out the latest models, I've realized that's not. That's. I. I don't believe that's the right approach, even if it were feasible. I came across a few days some really, really interesting models in the audio-to-audio task on Hugging Face. And one of them is specifically speaker isolation. And I did a double take when I saw this that someone's actually created a model not just for background noise removal, but for what you could call different speakers. So again, you're recording a business email to take an example, and you have friends over or you've got. Your wife is working from home. And the mechanism is something like listening to determining who the speaker is, getting a voice embedding from the first few seconds and then everyone that we recognize is a different voice we're just going to discard. So it's kind of acting at the VAD layer, so I'm going to just. It's going to take me at this point, there's so many clips, I think it's impossible to go back through it. I've definitely got clips of this background speech, but I'll try to just gather some intentionally because it's not so hard to do. And they can go into this one. So that's what this small subset is of the. Of the project.", "duration_s": 183.456, "sample_rate": 48000, "bit_depth": 16, "audio_format": "flac", "categories": ["background-speech", "speaker-isolation"], "target_terms": [], "prompt_id": "", "prompt_source": "drive-voice-note", "mode": "free-speech", "speaker": "Daniel Rosehill", "speaker_id": "8992", "speaker_gender": "male", "contributor": "Daniel Rosehill", "recorded_at": "2026-11-12T16:12:50Z", "device": "OnePlus Nord 3 5G", "microphone": "phone built-in", "environment": "home living room, others present", "transcript_status": "pending", "pronunciation": [], "task": "background-conversations", "noise_flags": ["background-speech", "baby-crying", "device-voice-prompt"], "model_hypothesis": "And recording a quick note here. And we're trying to do voice isolation. There is a man called Mike who's sitting on our couch. And this is a small, a very small data set that it's very. It's an experimentary one for people speaking in the background. So there's all kinds of people. So I'm just going to describe who I am. My name is Daniel, I'm the speaker here recording on the OnePlus Nord 3.5G and other people in the room. There's a man called Mike who's looking very unwell, lying on the couch. An Australian man. Ezra is my son and he's just triggered the roborock speaker. Hannah is my wife. And also made. It made a noise there. And we also had things. So this is a good first. First clip where we have. You can hear Ezra crying there. So this sample we have a challenge for background noise and for speaker isolation. We have non diarized audio. So we're getting a few angles for this one here. And the purpose of this data set, this is a very experimentary part of the project because one of the tasks that I wanted to train on for my ASR fine tune was specifically speaker isolation. And as I've mentioned in many of these blog posts, the challenges that I encounter dictating all the time are basically in the real world. It's not a pristine studio. People are having conversations. There's like PA messages. That's what you have when you're dictating. And I thought firstly, naively that there could be just record clips, ground truth is your speech. And therefore the model will eventually learn to discard everything ranging from bar music to what have you. And as, as I've been checking out the latest models, I've realized that's not. That's. I. I don't believe that's the right approach, even if it were feasible. I came across a few days some really, really interesting models in the audio to audio task on hugging face. And one of them is specifically speaker isolation. And I did a double take when I saw this that someone's actually created a model not just for background noise removal, but for what you could call different speakers. So again, you're recording a business email to take an example, and you have friends over or you've got. Your wife is working from home. And the mechanism is something like listening to determining who the speaker is, getting a voice embedding from the first few seconds and then everyone that we recognize is a different voice we're just going to discard. So it's kind of acting at the VAD layer, so I'm going to just. It's going to take me at this point, there's so many clips, I think it's impossible to go back through it. I've definitely got clips of this background speech, but I'll try to just gather some intentionally because it's not so hard to do. And they can go into this one. So that's what this small subset is of the. Of the project.", "asr_model": "assemblyai-universal (speaker_labels on; heard one speaker)", "notes": null, "extra": "{\"background_sources\": [{\"who\": \"Mike\", \"kind\": \"adult male speech\", \"note\": \"lying on the couch\"}, {\"who\": \"Ezra\", \"kind\": \"baby crying\"}, {\"who\": \"Hannah\", \"kind\": \"adult female speech\"}, {\"who\": \"Roborock\", \"kind\": \"device voice prompt\", \"note\": \"robot vacuum speaker, triggered by Ezra\"}], \"diarized\": false, \"source_file\": \"2026_11_12_18_12_50.opus\", \"transcript_note\": \"Foreground speech only (Daniel), verbatim. Draft = AssemblyAI with casing and names fixed by Claude (OnePlus Nord 3 5G, Roborock, Hugging Face, dataset); not yet checked against the audio by Daniel. Set transcript_status to reviewed once he has.\"}"}