Update to include entire dataset with metadata and models

Track .nemo files with Git LFS

Remove large files from history

Re-add models with LFS

.gitattributes

@@ -57,3 +57,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+*.nemo filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,20 +1,59 @@
 ---
 license: cc-by-sa-4.0
 ---
-# Frisian dialogue dataset: pre-release Dutch Speech Tech Day 2026
+# Frisian dialogue  (Project FRIS)
 
-NOTE: this is a pre-release of this dataset. A full version (adding around 5 more hours) will follow in March 2026.
+This dataset consists of transcribed spontaneous Frisian dialogue. The conversations are mocked consults between doctor and patient and their topics cover various medical concerns and ailments that are generally discussed in a GP practice. The dataset has been collected in a spontaneous conversational setting, resulting in realistic audio quality as encountered in real-world usage of speech-to-text models for medical scribes.
 
-This dataset consists of transcribed spontaneous Frisian dialogue. The conversations are (mocked) consults between doctor and patient and their topics cover various medical concerns and ailments that are generally discussed in a GP practice. 
+This repository contains the original audiofiles and transcripts as well as an aligned dataset ready for ASR training in the form of a NeMo manifest. This is divided into a training dataset of 24.0 hours and a test set of 2.6 hours.
 
-In this repo, you will find both the original audiofiles and transcripts as well as an aligned dataset ready for ASR training in the form of a NeMo manifest. This is divided into a training dataset of 22.54 hours and a corresponding test set of 2.54 hours.
+Number of speakers: 73
+
+Male to female ratio: 25:48
+
+### Files
+
+You can find the data structured as follows:
+- `manifest_train.json` the train split of the dataset as used to train the models
+- `manifest_test.json`  the test split of the dataset as used to evaluate the models 
+- `audio_aligned.tar.gz` the audio files the manifests refer to
+- `fris_raw.tar.gz` the raw data files: .mp4 track recordings of the conversations and the .docx transcriptions that correspond
+
+### Models
+
+The collection includes two NeMo models finetuned with this data from `nvidia/parakeet-tdt-0.6b-v3`:
+
+| Model name                  | Contents                      | CommonVoice eval (WER / CER) | FRIS eval (WER / CER)   | 
+|-----------------------------|-------------------------------|------------------------------|-------------------------|
+| `Juvoly-fyNL-FRIS`          | Project FRIS train            | 22.73% / 7.6%                | 17.5% / 9.1%            |
+| `Juvoly-fyNL-FRIS-withCV`   | Project FRIS train + CV train | 12.3% / 3.2%                 | 17.5% / 9.1%            |
+
+### Data structures
+
+The NeMo manifest contains the following data and metadata:
+- `audio_filepath` the path to the audio in the `audio_aligned` folder
+- `id`: the id of the audio fragment
+- `text`: the ground truth text corresponding to the audio
+- `duration`: duration of the audio
+- `segment_index`: segment index id
+- `word_count`: number of words
+- `src_offset`: offset in original audio file in seconds
+- `src_audio_filepath`: path to source audio
+- `speaker_role` being either Caregiver, Patient or Coordinator (a separate speaker doing announcement or coördination of the conversations)
+- `speaker_id` each speaker that participated was assigned a unique id
+- `gender` gender of the speaker
+- `dialect` the dialect of the speaker: KF (Kleifries), WF (Woudfries), NH (Noordhoeks), SWH (Zuidwesthoeks), Anders (other)
+- `age` the age group of the speaker (18-25, 26-35, 36-45, 46-55, 56-65, 65+)
+
+
+#### Example: 
+{"audio_filepath": "audio_aligned/Petear06 Migrêne_043_9.wav", "id": "Petear06 Migrêne_043_9", "text": "Uh, mar op it momint dat je sizze fan ja, ik ha faker dan fyftjin dagen yn 'e maand, ja, je pynstillers nedich, dan kinne je eigentlik ek wer troch dy pynstillers wer pineholle krije.", "duration": 11.7, "segment_index": 43, "word_count": 35, "src_audio_filepath": "Petear06 Migrêne.m4a", "speaker_role": "Caregiver", "speaker_id": 106, "gender": "F", "dialect": "WF", "src_offset": 536.5, "age": "36-45"}
+
+### Rights
 
 You are free to use this dataset both academically as well as commercially under the CC BY-SA license.
 
 The dataset has been curated as a collaboration between the Provinsje Fryslân, Taalstudio Prins and Juvoly.
 
-Questions: tessel@juvoly.nl
+Questions/contact: tessel@juvoly.nl
 
-Number of speakers: 73
-Male to female ratio: 25:48
-Dialects: 4 (Kleifries, Woudfries, Noordhoeks, Zuidwesthoeks)

test.tar.gz → audio_aligned.tar.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6ed24d6c3f5c72c2922e38a108ad811624006e5682582e6d630d2239df67525d
-size 248842692
+oid sha256:e64042294a1625dfeedda1e2b5b12e12d423667ab5b8e32a66841295cc14c2fa
+size 2587753355

fris-raw.tar.gz → fris_raw.tar.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6adb9c458dcc44620d35fd54a147ff6f5b73704d17aca59c83d3da0810423e38
-size 484270885
+oid sha256:95cf6e893803feb771d4cbbc742c51ddc5ad95825bbd56e5568b1998c89765c2
+size 550783126

train.tar.gz → models/Juvoly-fyNL-FRIS-CV.nemo

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cd1f7fbb38817075ac4e3e95d46b516b952436c5cc9dc3b73f5d483d0cc182d5
-size 1966815098
+oid sha256:1a5c32a8783a70b57e1dfc38378caace10c122ee97b9d2e4a19c448789e00cfb
+size 2511247360

models/Juvoly-fyNL-FRIS.nemo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5af497a1e0431994dc5217ff89dff2dbc737c1dbcd2c5cabcf5a9b52fdb7e22a
+size 2511247360