liva-ai/code-switching-asr

If the samples in the dataset viewer don't load, they can also be accessed here.

Code-Switching ASR

Code-switching ASR is a speech dataset of code-switching between English and medium to low resource languages. Samples include en-sw, en-pcm, en-yo, and en-tl but we can deliver any of the languages listed here: https://huggingface.co/datasets/liva-ai/yapdo-convo (the hours have yet to be updated as of 07/14/2026 as we have much higher volume now - and we can easily collect more of any language)

Samples are full verbatim with speaker diarization and audio tags (e.g. [laughing], [phone buzzing], etc.). Transcripts are either AI-generated and then human reviewed, or fully human generated, depending on customer request.

Audio

All audio is collected in-house from our own pool of human contributors. The audio is fully channel separated.

How We Collect Data

The difference is in how we collect the data. We collect it through our internal consumer platforms, where people socialize with friends, whereas other providers typically create environments where contractors are paid to speak with each other as strangers. We've learned that this makes a huge difference in the naturalness of the interactions.

Our apps have users around the world, which has allowed us to capture various low-resource languages, mixed languages (e.g. Sheng), and code-switching as they naturally occur in the wild, which are all areas where ASR models can improve greatly.

Technical Analysis

Property	Value
Sample rate	48 kHz
Bit depth	16-bit PCM
File format	WAV
Mean SNR	~33 dB
Median RMS	-26 dBFS
Average speech ratio	0.35
Spectral centroid	~0.66 kHz
Frequency content	3.3 kHz (averaged over 10-30 second clips)

Transcripts

All transcription is done in-house. Transcripts can follow a specific style guide (e.g. verbatim vs. non-verbatim, overlapping vs. non-overlapping timestamps, specific audio tags, etc.). We can provide transcripts for any of the languages listed here: https://huggingface.co/datasets/liva-ai/yapdo-convo

Transcription Process

All transcripts go through at least two full human review passes.

First pass: A native-speaker transcriber reviews and revises an AI-generated transcript or writes the transcript from scratch using our custom-built transcript editor. Transcribers are provided the conversation as channel-separated audio, which enables precise speaker diarization correction and makes it easier to identify and label each speaker accurately. They work through the entire audio file to produce a full verbatim transcript with precise timestamps down to the millisecond, labeled speaker turns for diarization, and audio event tags for non-speech sounds such as [phone buzzing], [laughing], or [door closing]. Native fluency allows transcribers to accurately capture code-switching, overlaps, disfluencies, colloquialisms, and other conversational details that are difficult to capture without deep familiarity with the speakers and language variety.

Second pass: Once the first pass is complete, the file is handed off to a senior reviewer with a proven track record of producing high-quality transcripts. The senior reviewer listens through the full audio file again and manually corrects any remaining transcription errors, spelling issues, timestamp inconsistencies, speaker label issues, or missed audio events. In addition, a senior native project lead performs rigorous spot checks across completed files to monitor quality, enforce consistency, and identify any recurring issues that need to be corrected across the project.

Transcriber quality: Transcribers are initially screened through a rigorous four-step process, involving both automation and human scoring, and are continuously monitored for attention and quality throughout every transcription project. Transcribers are promoted to more senior positions after delivering consistent quality work for weeks and continue to be monitored by Liva AI staff for attention to detail and quality.