Case Studies

The Challenge

LALAL.AI, a leading company in stem-splitting technology, faced a critical challenge in improving their AI and needed more Dutch input to improve model intonation patterns, lacking sufficient quality data for training.

The company needed:

• High-quality Dutch voice recordings with natural intonation
• Impeccable speech with no background noise
• Fast turnaround time to meet development deadlines
• Technical specifications that matched their existing dataset (-60dB SNR)

Our Solution

I delivered a short Dutch narrated script, audiobook-style. Focussed on expressive delivery with a 95 wpm pace, to improve prosodic cues like pauses and pitch shifts, supporting emotion and sentiment detection.

I used a professional-grade microphone and signal chain, with a sub -60db signal-to-noise ratio, to ensure sharp, detailed audio - providing the cleanest possible input for AI training.

Results

Client was very happy with the result - 0% error rate, 0 revisions needed. The data was used to contribute to more accurate voice cloning, complete with accurate emphases, intonations and nuances.

The Challenge

NIO's in-car assistant, NOMI, had robust English and Chinese speech recognition capabilities but lacked comprehensive Dutch language support. The company needed to expand NOMI's linguistic capabilities to serve the Dutch market effectively.

Key requirements included:

• Large volume of high-quality Dutch speech data
• Automotive-specific vocabulary and commands
• Clean audio suitable for in-car environments
• Consistent quality across all recordings

Our Solution

I recorded and delivered over 11,000 Dutch sentences, around 8 hours of clean voice data, for training NOMI. The project was completed in 6 days and didn’t require a single revision. The

• 8+ hours of diverse Dutch speech recordings
• Automotive-specific commands and conversational patterns
• Noise-free mobile recording to simulate car environment

Results

Better Dutch language recognition led to smoother interactions with NOMI — reducing user frustration and enhancing the driving experience for Dutch-speaking NIO customers.

The Challenge

Telus was conducting advanced research in Text-to-Speech (TTS) technology and needed natural-sounding Dutch male speech data to improve the expressiveness and accuracy of their TTS models. The existing synthetic voices lacked the natural intonation and emotional range required for their research objectives.

Specific requirements included:

• Natural, conversational Dutch voices
• Emotional range and expressiveness
• Unprocessed, lossless recordings for research purposes
• Diverse content covering various speech scenarios

Our Solution

I provided Telus with a specialized dataset focused on natural Dutch male speech patterns:

• Expressive scripted monologues with natural delivery
• Diverse content including conversational, narrative, and emotional speech
• Raw, high-fi recordings optimized for TTS research (-60dB SNR)
• Multiple speaking styles and emotional contexts

Results

Delivered over 3 hours of studio-quality narration in 4 days, boosting the clarity and realism of internal TTS outputs.