Hebrish: English-Hebrew code-switching
Collection
Every Hebrish model, dataset and Space, including those still on danielrosehill. hebrish.app • 35 items • Updated
Datasets:
id stringlengths 17 23 | file_name stringlengths 34 48 | date timestamp[s]date 2026-10-26 00:00:00 2026-10-31 00:00:00 | time stringlengths 8 8 | title stringlengths 7 39 | category stringclasses 5
values | tags listlengths 0 6 | speaker stringclasses 1
value | language stringclasses 1
value | transcript stringlengths 453 11.1k | for_claude bool 2
classes | claude_done_at timestamp[s]date 2026-10-31 00:00:00 2026-10-31 00:00:00 ⌀ | claude_reviewed_at timestamp[s]date 2026-10-31 00:00:00 2026-10-31 00:00:00 ⌀ | claude_decision stringclasses 1
value | claude_category stringclasses 4
values | claude_actioned bool 2
classes | claude_log stringlengths 193 633 ⌀ | duration_s float64 27.1 231 ⌀ | sample_rate int64 48k 48k ⌀ | bit_depth int64 24 24 ⌀ | audio_format stringclasses 1
value | source stringclasses 1
value | transcript_model stringclasses 2
values | recorded_at stringlengths 24 24 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2026-10-26-152336 | audio/2026-10-26_improvement-note-1_15-23-36.mp3 | 2026-10-26T00:00:00 | 15:23:36 | The Ideas dataset (meta-idea) | Dataset design | [
"ideas",
"meta-idea",
"dataset-design",
"capture-app"
] | Daniel Rosehill | en | This is a voice note I'm recording while I'm working on my voice note collection app for training data for my fine tunes, which is essentially whenever I have a spare moment, I'm just sending in some clips here and there. I've been trying to mix up recordings that were done at home, in different rooms, friends' places,... | false | null | null | null | null | false | null | null | null | null | null | null | null | null |
2026-10-26-153216 | audio/2026-10-26_improvement-note-2_15-32-16.mp3 | 2026-10-26T00:00:00 | 15:32:16 | Annotation design and the training plan | Annotation and training | [
"annotation",
"noise",
"background-conversation",
"background-music",
"training-plan",
"vocab-queue"
] | Daniel Rosehill | en | All right, so this is the second part of my notes for the dataset collection improvement that I was in the middle of there. If I go in some settings where I'm out and about and I'm collecting my voice notes, which are going to be tagged according to whether they're tech vocab, code-switching, as the case may be, probab... | false | null | null | null | null | false | null | null | null | null | null | null | null | null |
2026-10-26-155219 | audio/2026-10-26_improvement-note-3_15-52-19.mp3 | 2026-10-26T00:00:00 | 15:52:19 | Technical ideas to explore | Technical ideas | [
"noisy-speech",
"cross-lingual",
"speaker-verification",
"feasibility"
] | Daniel Rosehill | en | I have an idea maybe we can categorize these notes. I'm speaking to you, Claude Code, or I will be speaking to you, this is for your attention. We can categorize this dataset of ideas and decompose them into categories, and maybe each one can be decomposed into sections, but an idea came to me in the grocery store ther... | false | null | null | null | null | false | null | null | null | null | null | null | null | null |
20261030T104948Z_fhz3oh | audio/20261030T104948Z_fhz3oh.flac | 2026-10-30T00:00:00 | 10:49:48 | Room tone | Dataset design | [] | Daniel Rosehill | en | So, one quick thing to improve in my little capture app here is the room tone capture because it's not just room tone is useful, but I think background tone is more important. So, I'm going to capture this manually offline. 60 seconds of unfiltered, unprocessed of Jaffa Street as it is as I'm recording these this morni... | false | null | null | null | null | false | null | 230.373 | 48,000 | 24 | flac | capture-app | google/gemini-3.5-flash | 2026-10-30T10:49:48.196Z |
20261031T115507Z_e5j62q | audio/20261031T115507Z_e5j62q.flac | 2026-10-31T00:00:00 | 11:55:07 | Good samples | Other | [] | Daniel Rosehill | en | All right, this is just a note as I'm recording some samples. So, one that just came up for me was 900 Mbps, which of course one would naturally say 900 megabytes per second. Or in maybe more realistic internet speeds these days, 1.5 Gbps. I mean, some people would say Gbps, but I think more commonly people would say g... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | task-prompts | true | 2026-10-31: Reviewed: numerics already has 20 units prompts (incl. 900 Mbps); missing ms/latency, Gbps, MB/s, ounces, ml, tbsp, km/h. Plan ~60 prompts in data/task-prompts/numerics.json. Waiting on Daniel: abbreviated vs word targets (open question 37) and Mbps = megabits.
2026-10-31: Daniel settled: tech units abbrevi... | 97.125 | 48,000 | 24 | flac | capture-app | google/gemini-3.5-flash | 2026-10-31T11:55:07.542Z |
20261031T121138Z_0mj78n | audio/20261031T121138Z_0mj78n.flac | 2026-10-31T00:00:00 | 12:11:38 | Stop showing type | Other | [
"capture-app"
] | Daniel Rosehill | en | All right, so this is a note for the capture app. So when I'm doing the sentences and recording them, so right now there's a lot of Israeli trade names that I'm not familiar with and for these I'm going to have to like look them up or maybe get someone else to do it, like a doctor, if I can coerce someone or find someo... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | capture-app | true | 2026-10-31: Reviewed: wants a 'Stop showing <category>' choice under Skip (e.g. Israeli medical brand names), hidden until unhidden, with a Hidden list at the bottom of Tasks to restore. Not built yet; proposed to Daniel.
2026-10-31: Built in capture app release 109: Skip now offers 'Stop showing <type>' (bank categori... | 68.837 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:11:38.844Z |
20261031T122154Z_0vpjqq | audio/20261031T122154Z_0vpjqq.flac | 2026-10-31T00:00:00 | 12:21:54 | Word recurrence | Other | [] | Daniel Rosehill | en | Okay, so I'm recording at the moment and I feel like there are certain words that come up, so I've been doing samples with KDE Plasma in them today, and, you know, it's not a hugely important piece of vocab, I mean it's useful, but I question exactly how many samples we need and whether it's worth, especially as a larg... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | dataset | true | 2026-10-31: Built: ASR-Training data/term-importance.json rates all 5,515 bank target terms 1-5 (first pass by Claude subagents; anchors SSH key 5, KDE Plasma 3), mapped to 3/5/10/15/20 takes, unrated 3. Capture app release 109 serves a sentence last once every target term in it has reached its target; lexicon_saturati... | 107.237 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:21:54.233Z |
20261031T122710Z_8stpot | audio/20261031T122710Z_8stpot.flac | 2026-10-31T00:00:00 | 12:27:10 | Untitled note | Capture app | [] | Daniel Rosehill | en | Okay, so, one capture app fix here to do is in the idea notes, so we have subtabs for idea, background noise, and room tone. Let's just have like idea, background noise, room. I'm just trying to keep it like smaller, those tabs. And I don't need these descriptions: a voice note about the project, what to build next, be... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | capture-app | true | 2026-10-31: One media control bar shared by Idea, Background noise and Room (Pause, Stop, Retake | Record | Save, Delete); Recorder release 110, sprint planning/sprints/2026-10-31-ideas-review.md F3 | 60.661 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:27:10.701Z |
20261031T122817Z_mwo1c0 | audio/20261031T122817Z_mwo1c0.flac | 2026-10-31T00:00:00 | 12:28:17 | Improving | Capture app | [] | Daniel Rosehill | en | One other thing for the capture app, for the idea notes recording things, so it says after I, you know, stop it, it says that it's uploading. And maybe just have a little success thing or uploaded. And I feel like it should be polling more frequently because it was hanging there for like 20 seconds, which I feel probab... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | capture-app | true | 2026-10-31: Upload status listed under the bar; uploading hang fixed (pump released its lock before refreshing); category icons; shorter tab labels; Recorder release 110, sprint ideas-review F3 | 43.775 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:28:17.306Z |
20261031T122919Z_b6mny9 | audio/20261031T122919Z_b6mny9.flac | 2026-10-31T00:00:00 | 12:29:19 | Recording | Capture app | [] | Daniel Rosehill | en | And for this is also feedback on the um ideas tab in the recording app. So background noise room tone. Let's have a consistent media element throughout. Again same problem as before, and we just have one button. Let's standardize it to make it aligned with the sample recording in the um idea sorry in the main recording... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | capture-app | true | 2026-10-31: Slow transcription traced to default reasoning on Gemini Flash, sequential 60 s chunks and two commits per note; now one request with minimal reasoning, a 16 kbps Opus copy instead of WAV, one commit: 5-7 s for a 2 minute note. Stray last sentences came from a tiny tail chunk, removed. Recorder release 110,... | 27.119 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:29:19.555Z |
20261031T123221Z_sux6i5 | audio/20261031T123221Z_sux6i5.flac | 2026-10-31T00:00:00 | 12:32:21 | Dataset | Capture app | [] | Daniel Rosehill | en | So in the idea section of the app, I presume that my recordings are being uploaded immediately? We're not waiting on the general queue. That should be the case if not, and because these ones are, it's kind of a quick sync, and the other ones are in the background. Regarding the general push to using the Parquet format,... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | capture-app | true | 2026-10-31: Answer: Ideas notes upload immediately, independent of the take queue. Push by data size and slower child-dataset updates recorded as ideas, not actioned: planning/ideas/IDEA-0002.md, IDEA-0003.md | 120.234 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:32:21.865Z |
20261031T123640Z_d7zxpk | audio/20261031T123640Z_d7zxpk.flac | 2026-10-31T00:00:00 | 12:36:40 | Untitled note | Dataset design | [] | Daniel Rosehill | en | Okay, so today I was doing the readings and I came across Abarbanel Street, it's a bit of a tongue twister, Abarbanel Street in Jerusalem, which is a real street, and I was thinking that's actually a good one to have because definitely a use case here would be Waze, like navigation, where that's often actually quite sp... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | task-prompts | true | 2026-10-31: Israeli streets round 2: 42 streets named after people, from the national street list; 212 phone prompts after a review rejected 6. The most frequent names (Jabotinsky, Herzl and others) already had sentences and were left out. ASR-Training 21bd9ab, Recorder release 111, sprint ideas-review F6 | 179.333 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:36:40.271Z |
20261031T123810Z_g01hqv | audio/20261031T123810Z_g01hqv.flac | 2026-10-31T00:00:00 | 12:38:10 | Project organization | Other | [] | Daniel Rosehill | en | So, I'm sending in just a little batch of notes today, using the ideas tab which is coming along. And what I'd like to do, this is a message for the attention of Dear Claude. In fact, maybe that filter is bad. We can take it off because probably most things are going to be for Claude. At least it should be default turn... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | capture-app | true | 2026-10-31: For Claude ticked by default and after each save (Recorder release 111, F5). Skill claude-ideas now files every reviewed note into the Recorder planning tree: sprints for work done, features/to-do for backlog, ideas for not-yet (ASR-Training 21bd9ab, D1) | 77.274 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:38:10.893Z |
20261031T125226Z_2yjpzn | audio/20261031T125226Z_2yjpzn.flac | 2026-10-31T00:00:00 | 12:52:26 | Audio and | Other | [] | Daniel Rosehill | en | Here is another audio or a vocab set to review, which is basically common file names. So, this is one that I'll often see speech to text struggle with, especially when using Claude Code or it'll reach for very different versions every time. So on Kdenlive video editor, I think it's VAAPI, ROCm support, AMD. But the mor... | true | 2026-10-31T00:00:00 | 2026-10-31T00:00:00 | act | vocabulary | true | 2026-10-31: Tech pack media-formats: 35 terms (WebP, SVG, MP4, MKV, M4A, HEIC, VAAPI, ROCm, Kdenlive, transcode, downscale, 1080p, 16:9, 4:3, codecs), 154 sentences framed as instructions to an agent. Backticks for commands recorded as open question 51, not actioned. ASR-Training 21bd9ab, Recorder release 111, sprint i... | 230.661 | 48,000 | 24 | flac | capture-app | google/gemini-3.8-flash | 2026-10-31T12:52:26.652Z |
Hebrish Improvement Notes
Voice notes of ideas and incremental-improvement notes for the Hebrish fine-tune project and its recording app, captured in the moment and transcribed. Not training data: this dataset records what to improve and why, so the thinking behind the datasets, the annotation design and the training plans is kept in the open alongside the models.
Each row is one voice note: the original audio (audio/*.mp3) and its transcript, with a short
title, a category and tags. Later notes may supersede earlier ones — these are iterations.
Website: hebrish.app: get the Hebrish keyboard for Android and the desktop apps, see the results, and find every model, dataset and demo in one place.
Columns
| Column | Meaning |
|---|---|
id |
stable id (YYYY-MM-DD-HHMMSS, the recording time) |
file_name |
the audio file (MP3 for the first three notes, FLAC 48 kHz / 24-bit for notes recorded in the app) |
date, time |
when the note was recorded |
title |
short descriptive title |
category |
grouped subject: Dataset design · Annotation and training · Technical ideas |
tags |
free tags for the specific ideas in the note |
speaker, language |
speaker and spoken language of the recording |
transcript |
cleaned transcript (fillers and false starts removed, nothing paraphrased) |
for_claude |
true when Daniel flagged the note for Claude (Claude Code) to act on |
claude_done_at |
date Claude acted on a flagged note, else empty |
Notes so far
| Date | Title | Category |
|---|---|---|
| 2026-10-26 | The Ideas dataset (meta-idea) | Dataset design |
| 2026-10-26 | Annotation design and the training plan | Annotation and training |
| 2026-10-26 | Technical ideas to explore | Technical ideas |
Speaker: Daniel Rosehill, the dataset's author, who speaks English with an Irish accent and lives in Jerusalem. Every recording is his voice.
How this dataset is meant to grow
The plan (described in the first note) is for the recording app used to collect the training data to gain an Ideas page: voice notes recorded there are pushed to this dataset rather than into the training sets, uploaded by a service worker, and transcribed with a Gemini Flash model (OpenRouter), with the transcript written back into the row. Until that exists, notes are added by hand.
Written by Claude.
- Downloads last month
- 60