Datasets:

Hebrish
/

Hebrish-Ideas

@@ -2,3 +2,4 @@
 {"id":"2026-10-26-153216","file_name":"audio/2026-10-26_improvement-note-2_15-32-16.mp3","date":"2026-10-26","time":"15:32:16","title":"Annotation design and the training plan","category":"Annotation and training","tags":["annotation","noise","background-conversation","background-music","training-plan","vocab-queue"],"speaker":"Daniel Rosehill","language":"en","transcript":"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, probably within a fairly short amount of time, I will pick up naturally examples of the stuff that I annotate in my first iterations of this, which I will describe. These were just based upon the classic things you actually encounter when you're just doing dictation in the street: traffic noise, get a couple of honks there. That's the light rail train in Jerusalem. You will get there you go, little honk. Got extra value for money there.\n\nYou'll get a lot of honking, especially if you're near a junction. You'll get background music, background conversations. As we come into the winter, probably get samples with lots of wind noise on certain days. What I thought about was because I feel the benefit of this, my philosophy on annotation and this fine-tuning and training is that even if it's not immediately obvious why an annotation would be helpful, I'd rather have it, and if I don't drill down to that level of detail initially, I'd prefer it's there from the start.\n\nWhat I mean by that is background conversations. In Israel, we are going to hear Hebrew, Yiddish, Arabic, Russian, Amharic, and I'd like to have that as a multi-select option because you might have situations where multiple languages can be heard in the same sample. Then for background music, I thought also you could have one for level: low, medium, high, when someone's really blaring it, you're at a bar or whatever, that might be examples of high. Then background language music, because again, I'm thinking about this purely from the technical direction, what are the challenges an ASR model has. There's a busker. There's a busker in Jerusalem singing in English.\n\nI'm coming at this from the ASR level. An ASR model has somebody speaking. It's converting phonetics into it doesn't have an innate understanding that this is the speaker, doesn't identify as far as I know in real time, and that therefore this other person is not the speaker. The rationale for this annotation, and I explain my rationale in case it can be developed on further, is basically to tweak the composition of the fine-tune. I'll admit might be an interesting experiment. I don't think anyone's really gained this data. Background conversation, all of these I think should be multi-select, as I call them, enums, and nullable, because most clips won't have any of these artifacts present. I might be forgetting a few, but I'd rather have just a few main ones.\n\nWhat I'm thinking here is we haven't built yet an annotation UI where I can go in retrospectively and add my own annotations. So far it's just what I select in the UI is what's coming through raw. Occasionally I'll manually delete a take, but for the most part, it's been just all tagged as being indoors. We should create an annotation UI for me, playback sample by sample. I'm thinking as well, I don't know if this is useful, but it feels a pity not to do it, computed metrics, whatever potentially useful ones might be, I'm going to try sit down here for a sec, for this purpose, meaning whether it's my estimated WPM, any vocal characteristics that might be useful.\n\nFor example, when I'm recording out here, so we've got some Hebrew speech in the background, it's a noisy audio floor, very noisy, and I feel like that could be done as a backfill. In the annotation space, what I'd love to do, I don't want to listen through all of my so far hundreds of short sentences, that's not going to be a lot of fun. What I'd rather is a first pass could isolate the subset of the dataset that was outside of the standard distribution, it's unusually noisy, and I'm going to assume then that these ones probably I didn't update my tag. If they're very noisy, either we were having guests around or I just didn't update my UI, almost certainly the latter. Then I'd like to go into that subset and listen to them when I'm at home, and then I can just manually tag traffic sound present, honking artifacts present, background conversation, PA announcement.\n\nThen, this is my vision for the training plan, that will create a few golden samples, as they're called, and from that, the idea is that we can move to light supervision, so with those pairs, we can create a private classifier for the silver pairs that can then identify, okay, we've got a few samples of the Jerusalem light rail, and in the next iteration of the first-pass scanner, we can add a classifier for that defect, with the idea that over time the actual second annotation layer, the one that I'm not doing manually, can be trained, a trained model that actually ensures the integrity of the annotations, and then to the next stage of the layer.\n\nI'll probably send in some more notes over the coming days because my plan is, as a way to get some exercise and some fresh air, I'm just going to find random noisy spots. So long as the dataset mixture in the app is I want to have a good mixture so that if I sit down in a coffee shop and speak into my phone for 10 minutes, we're going to get some code-switching ones, we're going to naturally just encounter different things.\n\nOne other thing that I will just make note of quickly while I'm making notes, and we can add a skill for just keeping some notes in here or open thoughts, inevitably, in the course of collecting this data, especially trying to deliberately collect it in noisy places, I'm going to collect some copyrighted music as background. I hope this would be protected under fair use. The only reason I flag it is maybe if it would affect the dataset, could be flagged in some copyright scanners that we might have to add a scrub for that, but then we'd lose I think a very, very useful training task in this fine-tune, which is increasing the robustness of ASR models to pick up speaker audio when there is music in the background.\n\nThose are in my day-to-day use of this technology for two years, gosh, I'm terribly unfit. I've become terribly unfit; I was very fit till recently. In my day-to-day use of this tech, that's my list of this is the 5% gap between, \"Yeah, this is fantastic, I can use it and just get polished emails,\" and it's almost there. It's the code-switching, lack of vocab, it's failing on the tech vocab, and it heard some random words that a street worker shouted in Arabic and thought that that was intended, and rendered it in random English stuff, and so the transcript wound up with whatever that might have sounded like in English. Those are all the tasks, and this one is a bit longer than I thought.\n\nWhat I want to do, in fact, disfluency removal is a great task. What I want to do, this is terrible, I'm categorizing my own disfluencies as I try to say this, I just did a false start. What I'd like to do is I think a few fine-tunes. The one I'm focusing on at the moment is Sabikh for the simple reason of it's the one that's most useful to me. While I'm delighted to open source this, my main motivation for wanting to do this is that I use ASR tools three hours per day is very conservative, I use this all the time. I have a strong vested interest in improving accuracy for my own day-to-day productivity, and that composite model is the one that's going to be most useful for me. Over time, when I get time to do this just for the benefit of other folks, I want to train up new checkpoints of the tech one and the combined ASR post-processing single forward pass model, which is a really much more experimentary design than this one.\n\nAnyway, these are a couple of long-form audio notes for anyone wondering if a person reading weird sentences about Cnaidlech and gosh, there have been some classics in the dataset. I'm trying very hard not to laugh often, and then I worry that I'm sounding like a robot because I'm suppressing I have some screenshots of the most ridiculous sentences that the thing has thrown up so far. I'm going to try inject a little bit more life into my recordings, to do prosody, because real people, when they're doing voice-to-text, I guess, they do all sorts of things.\n\nThe final thing, this is just very free-form at this point now, is that I'm trying to find on Hugging Face a lot of these artificial sentences, the ones done in haiku batches, maybe they're going to add a little bit of a nudge to the accent recognition and to the general great cause of reducing word error rate, but they're so boring, they're so lifeless, they're never things I would say. I'm wondering there must be datasets of the ones that would be useful, I was thinking about this today, if I could find one on Hugging Face for emails, calendar appointments, and I'd say one I'm constantly using voice-to-text to send in prompts. What I would need is them cleaned, so if the datasets are ASR transcripts, I need them normalized, without the disfluency, so I can actually read a clean reference text. But at least that might get me to, \"Oh, this is actually something I do, this is the kind of thing I actually use every day,\" as opposed to just, \"These are completely random sentences about anything else.\"\n\nOther random note and idea is I like the idea of, for this project, creating a podcast. I already created an XML feed based on my first sort of description of the project, an introduction to it, what my objectives were, why I want to do this, what the dataset design is, blah, blah, blah, all that stuff. I think it was an hour or so, I recorded that last week. I like the idea of actually using some of these updates to create ones, but these ones are much, much rougher, this one is just me literally brain dumping a bunch of ideas.\n\nI just said the word \"brain dumping\", and that's maybe one more thing that I need to work on, or an idea for iterative improvement here. I think as this goes on, we have one kind of self-correction loop, which is a Claude skill I have that I'm using my own transcription history, and I'm saying, \"Claude, take a look at the last I've been transcribing for four hours, export my history, and can you figure out stuff that the model has got wrong, and I'll train on the corrected ones?\" That works nicely. Probably the same thing can be done from the phone, but the parallel process I need is some easy way of when I'm not at the computer, maybe it's a Telegram bot, some easy service, maybe just a button in my recording UI of queuing up vocab for the list. For example, \"brain dumping\", right, that's exactly the kind of weird word that I don't say that often, but it's very unlikely that that has appeared in the lexicon so far. Maybe \"brain dump\" was in there, probably not. So great, two words to add to pad out the long tail of just random words. I want to add that as well. That's a queued task for the UI. I think we've got a few now. All right, that's improvement log number two for the app today."}
 {"id":"2026-10-26-155219","file_name":"audio/2026-10-26_improvement-note-3_15-52-19.mp3","date":"2026-10-26","time":"15:52:19","title":"Technical ideas to explore","category":"Technical ideas","tags":["noisy-speech","cross-lingual","speaker-verification","feasibility"],"speaker":"Daniel Rosehill","language":"en","transcript":"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 there.\n\nThere's a couple of tasks that I think are very important beyond the vocab stuff. The one I think that I'd be most interested to see the results for is training on noisy samples with background conversations and music. I'm particularly interested because of the cross-linguistic element. There was a loud conversation in Russian versus there was a loud conversation going on in English. If the gradient between me and them is the same, is it harder to, do we see any unintended rendering of the same language speech or different language speech? I'd be surprised if anyone's done data on that, but that's exactly the kind of thing that would make a big difference for me. You're commuting, you're recording notes, you're dictating, people are passing by having random conversations.\n\nMy question for that is really it's a technical one again. Is it possible to train on that? If it's trained on just through a crude process of sample training, would that actually be strong enough to eventually bake in the pattern by which the model, it's not a sophisticated heuristic for the model to say, it doesn't have that innate concept of this is the speaker, this is another voice, so it would be purely based upon a volume gradient, a signal gradient.\n\nThen I was thinking, wait, what if you had, take this implementation as an example, you have your dictation app running this model, and the user records a 60-second speech sample, a reference audio. From that 60 or 30 seconds or three minutes, whatever the case may be, we compute embeddings of the voice profile of the person. The question is if a model like this, which is intended for high-performance, very high-accuracy ASR for a user, in the inference stage, could say, as it processes audio tokens, checking whether those tokens match the embeddings, if there's a close match. If so, it passes; if no, in other words, there's like a gate within the inference layer that if there's a certain degree of separation between the known speaker voice and the other voice, then those tokens don't get processed. In other words, an actual technical implementation around that. Of course, defense in depth is the right way to do it: try the first and then maybe see the second.\n\nThat's just an idea that I wanted to throw out. We could categorize this one under technical ideas to explore. It's just a random idea that really needs feasibility research into the actual internals of this model and whether, always infer in prior art. This is probably something I need to bake in as what I call a habit of Claude these days. It's my term for always-assumed preferences. Whenever I ask Claude to see if something is possible, it can always infer that I want to see if other people have done it, because that's often the easiest way: look if anyone has made this; if no, is it even feasible? Usually you get the answer from the two. So that's the idea here to look into."}
 {"id":"20261030T104948Z_fhz3oh","file_name":"audio/20261030T104948Z_fhz3oh.flac","date":"2026-10-30","time":"10:49:48","title":"Room tone","category":"Dataset design","tags":[],"speaker":"Daniel Rosehill","language":"en","transcript":"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 morning samples, which isn't terribly loud as Jerusalem goes, actually. It's a pretty chill afternoon here. Lovely, mild, late, well, is it late morning? Early afternoon, I guess, technically, as people are, you know, the few hours before Shabbat. Other side of Jaffa by the Shuk on Agripas is much, much, you know, louder and this. So, this is actually a perfect background noise to capture because you've got that kind of low-level chatter in the background, little the odd beep, the odd train, but nothing too like kind of, I guess, you know, none of the classic. Okay, a guy just honked his horn three feet from me kind of thing. So I'm going to just record 60 seconds and just as a note to self, I discovered this week for recording unprocessed on Android, I'm a big fan of an app called ASR Mic. It's a terrific and the best voice recorder on Android. And we have this noise in our apartment coming from somewhere nearby. I sound like the biggest noise crank recording this. It's not really my personality at all. But we have just very specific bar driving us a little bit up the walls. And it's opened like a month ago and they're just playing this like bass that just gets into our bedroom at like 10:00, 11:00, midnight, 1:00 in the morning. So, I was of course using Claude and I said, \"Claude, we're going to figure this out, man.\" Man or woman, you know, Claude is not gendered. And I said, \"What should I record? Can you figure out the frequency if I record for a clip. And it said, \"Sure, but make sure you record in raw and unprocessed.\" So, I didn't even know that there was such a thing on Android that you can just record raw. And so I did, and it figured out the frequency range, which is pretty cool. So, anyway, it was available in ASR Mic, and I'm going to just do the unprocessed recording here. So, we get 60 seconds. That was the Jerusalem train, light rail going by. Speaking of things in my dataset I need vocab for, Kfir, which is the both a boy's name and the esteemed operator of the light rail in Jerusalem. I don't think I have it yet. Am I going crazy reading these sentences? Only slightly. I need to add a bit more variety. I keep thinking of there has to be a way to gamify this because I keep noticing like between these recording sessions, when I do the training, it's very motivating because the accuracy goes up, and then I install it on my other device, and then I'm like, \"Hey, this thing's actually getting better. This is working.\" And then, if for anyone who Whoever listens to this note or looks at the data, there's no rhyme or reason, I'm just like, I'll do 30 minutes now. Last night I recorded while I was a little bit tipsy. We don't have an annotation yet for estimated blood alcohol level, but if you record all the time and you record voice notes, you're going to record voice notes when you're sober and you're going to record voice notes when you're not sober in bars, that's just how it is. So, that might be a useful tag. I wasn't wasted. I was a couple of beers and couple of glasses of wine in, but I did go out of my way to record a few in such mental conditions and in such acoustic conditions before deciding that I've done enough speaking to my phone for this week, I'm going to just enjoy the night. But anyway, got a few of those and that's it for today's voice capture app.","duration_s":230.373,"sample_rate":48000,"bit_depth":24,"audio_format":"flac","source":"capture-app","transcript_model":"google/gemini-3.5-flash","recorded_at":"2026-10-30T10:49:48.196Z"}

 {"id":"2026-10-26-153216","file_name":"audio/2026-10-26_improvement-note-2_15-32-16.mp3","date":"2026-10-26","time":"15:32:16","title":"Annotation design and the training plan","category":"Annotation and training","tags":["annotation","noise","background-conversation","background-music","training-plan","vocab-queue"],"speaker":"Daniel Rosehill","language":"en","transcript":"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, probably within a fairly short amount of time, I will pick up naturally examples of the stuff that I annotate in my first iterations of this, which I will describe. These were just based upon the classic things you actually encounter when you're just doing dictation in the street: traffic noise, get a couple of honks there. That's the light rail train in Jerusalem. You will get there you go, little honk. Got extra value for money there.\n\nYou'll get a lot of honking, especially if you're near a junction. You'll get background music, background conversations. As we come into the winter, probably get samples with lots of wind noise on certain days. What I thought about was because I feel the benefit of this, my philosophy on annotation and this fine-tuning and training is that even if it's not immediately obvious why an annotation would be helpful, I'd rather have it, and if I don't drill down to that level of detail initially, I'd prefer it's there from the start.\n\nWhat I mean by that is background conversations. In Israel, we are going to hear Hebrew, Yiddish, Arabic, Russian, Amharic, and I'd like to have that as a multi-select option because you might have situations where multiple languages can be heard in the same sample. Then for background music, I thought also you could have one for level: low, medium, high, when someone's really blaring it, you're at a bar or whatever, that might be examples of high. Then background language music, because again, I'm thinking about this purely from the technical direction, what are the challenges an ASR model has. There's a busker. There's a busker in Jerusalem singing in English.\n\nI'm coming at this from the ASR level. An ASR model has somebody speaking. It's converting phonetics into it doesn't have an innate understanding that this is the speaker, doesn't identify as far as I know in real time, and that therefore this other person is not the speaker. The rationale for this annotation, and I explain my rationale in case it can be developed on further, is basically to tweak the composition of the fine-tune. I'll admit might be an interesting experiment. I don't think anyone's really gained this data. Background conversation, all of these I think should be multi-select, as I call them, enums, and nullable, because most clips won't have any of these artifacts present. I might be forgetting a few, but I'd rather have just a few main ones.\n\nWhat I'm thinking here is we haven't built yet an annotation UI where I can go in retrospectively and add my own annotations. So far it's just what I select in the UI is what's coming through raw. Occasionally I'll manually delete a take, but for the most part, it's been just all tagged as being indoors. We should create an annotation UI for me, playback sample by sample. I'm thinking as well, I don't know if this is useful, but it feels a pity not to do it, computed metrics, whatever potentially useful ones might be, I'm going to try sit down here for a sec, for this purpose, meaning whether it's my estimated WPM, any vocal characteristics that might be useful.\n\nFor example, when I'm recording out here, so we've got some Hebrew speech in the background, it's a noisy audio floor, very noisy, and I feel like that could be done as a backfill. In the annotation space, what I'd love to do, I don't want to listen through all of my so far hundreds of short sentences, that's not going to be a lot of fun. What I'd rather is a first pass could isolate the subset of the dataset that was outside of the standard distribution, it's unusually noisy, and I'm going to assume then that these ones probably I didn't update my tag. If they're very noisy, either we were having guests around or I just didn't update my UI, almost certainly the latter. Then I'd like to go into that subset and listen to them when I'm at home, and then I can just manually tag traffic sound present, honking artifacts present, background conversation, PA announcement.\n\nThen, this is my vision for the training plan, that will create a few golden samples, as they're called, and from that, the idea is that we can move to light supervision, so with those pairs, we can create a private classifier for the silver pairs that can then identify, okay, we've got a few samples of the Jerusalem light rail, and in the next iteration of the first-pass scanner, we can add a classifier for that defect, with the idea that over time the actual second annotation layer, the one that I'm not doing manually, can be trained, a trained model that actually ensures the integrity of the annotations, and then to the next stage of the layer.\n\nI'll probably send in some more notes over the coming days because my plan is, as a way to get some exercise and some fresh air, I'm just going to find random noisy spots. So long as the dataset mixture in the app is I want to have a good mixture so that if I sit down in a coffee shop and speak into my phone for 10 minutes, we're going to get some code-switching ones, we're going to naturally just encounter different things.\n\nOne other thing that I will just make note of quickly while I'm making notes, and we can add a skill for just keeping some notes in here or open thoughts, inevitably, in the course of collecting this data, especially trying to deliberately collect it in noisy places, I'm going to collect some copyrighted music as background. I hope this would be protected under fair use. The only reason I flag it is maybe if it would affect the dataset, could be flagged in some copyright scanners that we might have to add a scrub for that, but then we'd lose I think a very, very useful training task in this fine-tune, which is increasing the robustness of ASR models to pick up speaker audio when there is music in the background.\n\nThose are in my day-to-day use of this technology for two years, gosh, I'm terribly unfit. I've become terribly unfit; I was very fit till recently. In my day-to-day use of this tech, that's my list of this is the 5% gap between, \"Yeah, this is fantastic, I can use it and just get polished emails,\" and it's almost there. It's the code-switching, lack of vocab, it's failing on the tech vocab, and it heard some random words that a street worker shouted in Arabic and thought that that was intended, and rendered it in random English stuff, and so the transcript wound up with whatever that might have sounded like in English. Those are all the tasks, and this one is a bit longer than I thought.\n\nWhat I want to do, in fact, disfluency removal is a great task. What I want to do, this is terrible, I'm categorizing my own disfluencies as I try to say this, I just did a false start. What I'd like to do is I think a few fine-tunes. The one I'm focusing on at the moment is Sabikh for the simple reason of it's the one that's most useful to me. While I'm delighted to open source this, my main motivation for wanting to do this is that I use ASR tools three hours per day is very conservative, I use this all the time. I have a strong vested interest in improving accuracy for my own day-to-day productivity, and that composite model is the one that's going to be most useful for me. Over time, when I get time to do this just for the benefit of other folks, I want to train up new checkpoints of the tech one and the combined ASR post-processing single forward pass model, which is a really much more experimentary design than this one.\n\nAnyway, these are a couple of long-form audio notes for anyone wondering if a person reading weird sentences about Cnaidlech and gosh, there have been some classics in the dataset. I'm trying very hard not to laugh often, and then I worry that I'm sounding like a robot because I'm suppressing I have some screenshots of the most ridiculous sentences that the thing has thrown up so far. I'm going to try inject a little bit more life into my recordings, to do prosody, because real people, when they're doing voice-to-text, I guess, they do all sorts of things.\n\nThe final thing, this is just very free-form at this point now, is that I'm trying to find on Hugging Face a lot of these artificial sentences, the ones done in haiku batches, maybe they're going to add a little bit of a nudge to the accent recognition and to the general great cause of reducing word error rate, but they're so boring, they're so lifeless, they're never things I would say. I'm wondering there must be datasets of the ones that would be useful, I was thinking about this today, if I could find one on Hugging Face for emails, calendar appointments, and I'd say one I'm constantly using voice-to-text to send in prompts. What I would need is them cleaned, so if the datasets are ASR transcripts, I need them normalized, without the disfluency, so I can actually read a clean reference text. But at least that might get me to, \"Oh, this is actually something I do, this is the kind of thing I actually use every day,\" as opposed to just, \"These are completely random sentences about anything else.\"\n\nOther random note and idea is I like the idea of, for this project, creating a podcast. I already created an XML feed based on my first sort of description of the project, an introduction to it, what my objectives were, why I want to do this, what the dataset design is, blah, blah, blah, all that stuff. I think it was an hour or so, I recorded that last week. I like the idea of actually using some of these updates to create ones, but these ones are much, much rougher, this one is just me literally brain dumping a bunch of ideas.\n\nI just said the word \"brain dumping\", and that's maybe one more thing that I need to work on, or an idea for iterative improvement here. I think as this goes on, we have one kind of self-correction loop, which is a Claude skill I have that I'm using my own transcription history, and I'm saying, \"Claude, take a look at the last I've been transcribing for four hours, export my history, and can you figure out stuff that the model has got wrong, and I'll train on the corrected ones?\" That works nicely. Probably the same thing can be done from the phone, but the parallel process I need is some easy way of when I'm not at the computer, maybe it's a Telegram bot, some easy service, maybe just a button in my recording UI of queuing up vocab for the list. For example, \"brain dumping\", right, that's exactly the kind of weird word that I don't say that often, but it's very unlikely that that has appeared in the lexicon so far. Maybe \"brain dump\" was in there, probably not. So great, two words to add to pad out the long tail of just random words. I want to add that as well. That's a queued task for the UI. I think we've got a few now. All right, that's improvement log number two for the app today."}
 {"id":"2026-10-26-155219","file_name":"audio/2026-10-26_improvement-note-3_15-52-19.mp3","date":"2026-10-26","time":"15:52:19","title":"Technical ideas to explore","category":"Technical ideas","tags":["noisy-speech","cross-lingual","speaker-verification","feasibility"],"speaker":"Daniel Rosehill","language":"en","transcript":"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 there.\n\nThere's a couple of tasks that I think are very important beyond the vocab stuff. The one I think that I'd be most interested to see the results for is training on noisy samples with background conversations and music. I'm particularly interested because of the cross-linguistic element. There was a loud conversation in Russian versus there was a loud conversation going on in English. If the gradient between me and them is the same, is it harder to, do we see any unintended rendering of the same language speech or different language speech? I'd be surprised if anyone's done data on that, but that's exactly the kind of thing that would make a big difference for me. You're commuting, you're recording notes, you're dictating, people are passing by having random conversations.\n\nMy question for that is really it's a technical one again. Is it possible to train on that? If it's trained on just through a crude process of sample training, would that actually be strong enough to eventually bake in the pattern by which the model, it's not a sophisticated heuristic for the model to say, it doesn't have that innate concept of this is the speaker, this is another voice, so it would be purely based upon a volume gradient, a signal gradient.\n\nThen I was thinking, wait, what if you had, take this implementation as an example, you have your dictation app running this model, and the user records a 60-second speech sample, a reference audio. From that 60 or 30 seconds or three minutes, whatever the case may be, we compute embeddings of the voice profile of the person. The question is if a model like this, which is intended for high-performance, very high-accuracy ASR for a user, in the inference stage, could say, as it processes audio tokens, checking whether those tokens match the embeddings, if there's a close match. If so, it passes; if no, in other words, there's like a gate within the inference layer that if there's a certain degree of separation between the known speaker voice and the other voice, then those tokens don't get processed. In other words, an actual technical implementation around that. Of course, defense in depth is the right way to do it: try the first and then maybe see the second.\n\nThat's just an idea that I wanted to throw out. We could categorize this one under technical ideas to explore. It's just a random idea that really needs feasibility research into the actual internals of this model and whether, always infer in prior art. This is probably something I need to bake in as what I call a habit of Claude these days. It's my term for always-assumed preferences. Whenever I ask Claude to see if something is possible, it can always infer that I want to see if other people have done it, because that's often the easiest way: look if anyone has made this; if no, is it even feasible? Usually you get the answer from the two. So that's the idea here to look into."}
 {"id":"20261030T104948Z_fhz3oh","file_name":"audio/20261030T104948Z_fhz3oh.flac","date":"2026-10-30","time":"10:49:48","title":"Room tone","category":"Dataset design","tags":[],"speaker":"Daniel Rosehill","language":"en","transcript":"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 morning samples, which isn't terribly loud as Jerusalem goes, actually. It's a pretty chill afternoon here. Lovely, mild, late, well, is it late morning? Early afternoon, I guess, technically, as people are, you know, the few hours before Shabbat. Other side of Jaffa by the Shuk on Agripas is much, much, you know, louder and this. So, this is actually a perfect background noise to capture because you've got that kind of low-level chatter in the background, little the odd beep, the odd train, but nothing too like kind of, I guess, you know, none of the classic. Okay, a guy just honked his horn three feet from me kind of thing. So I'm going to just record 60 seconds and just as a note to self, I discovered this week for recording unprocessed on Android, I'm a big fan of an app called ASR Mic. It's a terrific and the best voice recorder on Android. And we have this noise in our apartment coming from somewhere nearby. I sound like the biggest noise crank recording this. It's not really my personality at all. But we have just very specific bar driving us a little bit up the walls. And it's opened like a month ago and they're just playing this like bass that just gets into our bedroom at like 10:00, 11:00, midnight, 1:00 in the morning. So, I was of course using Claude and I said, \"Claude, we're going to figure this out, man.\" Man or woman, you know, Claude is not gendered. And I said, \"What should I record? Can you figure out the frequency if I record for a clip. And it said, \"Sure, but make sure you record in raw and unprocessed.\" So, I didn't even know that there was such a thing on Android that you can just record raw. And so I did, and it figured out the frequency range, which is pretty cool. So, anyway, it was available in ASR Mic, and I'm going to just do the unprocessed recording here. So, we get 60 seconds. That was the Jerusalem train, light rail going by. Speaking of things in my dataset I need vocab for, Kfir, which is the both a boy's name and the esteemed operator of the light rail in Jerusalem. I don't think I have it yet. Am I going crazy reading these sentences? Only slightly. I need to add a bit more variety. I keep thinking of there has to be a way to gamify this because I keep noticing like between these recording sessions, when I do the training, it's very motivating because the accuracy goes up, and then I install it on my other device, and then I'm like, \"Hey, this thing's actually getting better. This is working.\" And then, if for anyone who Whoever listens to this note or looks at the data, there's no rhyme or reason, I'm just like, I'll do 30 minutes now. Last night I recorded while I was a little bit tipsy. We don't have an annotation yet for estimated blood alcohol level, but if you record all the time and you record voice notes, you're going to record voice notes when you're sober and you're going to record voice notes when you're not sober in bars, that's just how it is. So, that might be a useful tag. I wasn't wasted. I was a couple of beers and couple of glasses of wine in, but I did go out of my way to record a few in such mental conditions and in such acoustic conditions before deciding that I've done enough speaking to my phone for this week, I'm going to just enjoy the night. But anyway, got a few of those and that's it for today's voice capture app.","duration_s":230.373,"sample_rate":48000,"bit_depth":24,"audio_format":"flac","source":"capture-app","transcript_model":"google/gemini-3.5-flash","recorded_at":"2026-10-30T10:49:48.196Z"}
+{"id":"20261031T115507Z_e5j62q","file_name":"audio/20261031T115507Z_e5j62q.flac","date":"2026-10-31","time":"11:55:07","title":"Good samples","category":"Other","tags":[],"speaker":"Daniel Rosehill","language":"en","transcript":null,"duration_s":97.125,"sample_rate":48000,"bit_depth":24,"audio_format":"flac","source":"capture-app","transcript_model":null,"recorded_at":"2026-10-31T11:55:07.542Z"}