Update README.md

README.md

@@ -32,12 +32,12 @@ size_categories:
 
 **Utterly** is a speech dataset derived from *pipecat-ai/human_5_all*. It contains **~3.86k English utterances** by a broad range of speakers, and augments conversational audio with turn-level annotations, including:
 
-* Audio utterances
-* Whisper-generated transcripts
-* Speaker identifiers (TODO)
+* Verbatim whisper-generated transcripts
 * End-of-turn (EoT) markers
+* Speaker identifiers (Coming soon)
 
-The dataset is designed to support research and development of speech and dialogue systems that require joint modeling of **speech recognition and conversational turn-taking**, such as streaming ASR systems and real-time conversational agents.
+
+The dataset is designed to support research and development of speech and dialogue systems that require joint modeling of **speech recognition** and **conversational turn-taking**, such as streaming ASR systems and real-time conversational agents.
 
 ---
 
@@ -47,8 +47,12 @@ The dataset is designed to support research and development of speech and dialog
 * **Language(s)**: English
 * **Modality**: Audio (speech; mono-channel; sampled at 16kHz), Text
 * **Interaction type**: Human conversational speech
+* **Utterances**: 3,860
+* **Speakers**: 100+
+
+Dataset splits (e.g., train/validation/test) are not predefined and may be created by downstream users as needed.
 
-The Utterly dataset is a *derived dataset*. All audio originates from the base dataset, with additional annotations created by the dataset author.
+Note that Utterly is a *derived dataset*. All audio originates from the base dataset, with additional annotations created by the dataset author.
 
 ---
 
@@ -57,8 +61,7 @@ The Utterly dataset is a *derived dataset*. All audio originates from the base d
 * **Transcripts**
 
   * Generated automatically using **Whisper Large V3 Turbo**.
-  * A subset of samples was manually reviewed and corrected.
-  * Based on this manually corrected subset, the transcripts are estimated to have approximately **98.3% transcription accuracy**, corresponding to a very low word error rate.
+  * A subset of samples (~200) was manually reviewed and corrected. The transcripts are estimated to have approximately a word error rate (WER) of **~2.8%**.
 
 * **End-of-Turn markers**
 
@@ -66,7 +69,7 @@ The Utterly dataset is a *derived dataset*. All audio originates from the base d
 
 * **Speaker IDs**
 
-  * TODO
+  * Coming soon
 
 ---
 
@@ -76,7 +79,7 @@ A typical data entry includes:
 
 * `audio`: Path or reference to the audio utterance
 * `transcript`: Text transcription of the utterance
-* `speaker_id`: Identifier for the speaker (TODO)
+* `speaker_id`: Identifier for the speaker (Coming soon)
 * `is_completed`: Boolean or categorical flag indicating end-of-turn, i.e. turn completion
 
 Depending on downstream usage, additional metadata from the source dataset may also be present.
@@ -95,17 +98,6 @@ The Utterly dataset is designed to support a range of speech and dialogue resear
 
 ---
 
-## Dataset Size and Composition
-
-* **Number of utterances**: 3,860
-* **Language**: English
-* **Modalities**: Audio, text
-* **Speakers**: Multiple (as defined in the source dataset)
-
-Dataset splits (e.g., train/validation/test) are not predefined and may be created by downstream users as needed.
-
----
-
 ## Quality Considerations
 
 * End-of-turn annotations involve human judgment and may reflect subjective interpretations of conversational completion.