Voice Assistants in Singapore Homes: Real-World Performance

Voice assistant adoption in Singapore is high relative to regional neighbours, partly due to widespread broadband penetration and partly because controlling air conditioning by voice has obvious appeal in a climate where the AC runs year-round. This article documents what three assistants — Amazon Alexa, Google Assistant, and Apple Siri — actually deliver when placed in Singapore households over a three-month period.

Amazon Echo Dot 3rd Gen — one of the most widely used smart speakers in Singapore homes, primarily for Alexa-based smart home control.

The Singapore English Variable

Singlish — the creole spoken across Singapore's ethnic communities — presents a particular challenge for voice recognition models trained predominantly on American, British, and Australian English corpora. Command accuracy testing was conducted with five participants: one speaking predominantly standard Singapore English, two Mandarin-accented English speakers, one Tamil-accented English speaker, and one Malay-accented English speaker.

Commands were standardised across 30 phrases per session covering smart home control (lights, AC, appliances), information queries (local bus timings, prayer times, MRT disruption status), and media control. Each command was issued three times per participant to control for ambient noise variation.

Recognition Accuracy by Assistant

Across all participants and commands, Google Assistant achieved the highest overall recognition rate at 91.4%. This aligns with Google's documented investment in Southeast Asian language model training — the company has published research specifically on Singapore English phonological patterns. Alexa followed at 84.7%, with most misrecognitions occurring in commands with Mandarin-influenced vowel shifts. Siri came in at 79.2%, with notable difficulty on Tamil-accented commands involving dental consonants.

These numbers tell one part of the story. What the numbers don't capture is the distribution of failures: Google's misrecognitions were more evenly spread and resulted in "I didn't understand that" responses. Alexa's failures more frequently produced confident but wrong responses — executing the wrong command rather than acknowledging the ambiguity. For home automation, an unexecuted command is preferable to a misdirected one.

Local Query Handling

Voice assistants struggle with hyper-local Singapore queries in ways that are not immediately apparent to users. Some examples from the testing period:

• "What time does the next 65 bus arrive at Tampines Interchange?" — Google Assistant correctly accessed SBS Transit data via its integration. Alexa returned a generic response about bus schedules. Siri said it could not help with that request.
• "Is the Circle Line running normally?" — Google Assistant pulled a current status summary. Alexa directed to a generic transit website. Siri provided no relevant response.
• "What's the PSI right now?" — All three assistants failed this. None connected to NEA's air quality data despite it being publicly available. PSI queries are a recurrent frustration during haze season from Sumatra burning.

Air Conditioning Control: The Most Relevant Use Case

In a Singapore household with AC running most hours of the day, voice control of cooling is the most frequently used smart home function. All three ecosystems support AC control, but the path differs:

Google Assistant integrates directly with Google Home and supports a wide range of AC brands via Matter and legacy Works with Google protocols. Daikin, Mitsubishi, and Panasonic — the three dominant brands in Singapore residential installations — all have official integrations. Temperature commands ("set AC to 24 degrees") execute reliably. Mode changes (fan speed, dry mode) work with some brands but not others depending on the model generation.

Alexa handles AC control through third-party skills, most commonly the Sensibo skill if a Sensibo IR blaster is in use. The Sensibo-Alexa integration is stable and covers mode switching, temperature, and scheduling. Without a hub like Sensibo, controlling an older split-unit AC via Alexa requires an IR blaster setup that Alexa does not manage natively.

Siri requires HomeKit compatibility or a HomeKit bridge. Most Singapore-market AC units do not ship with HomeKit support. Achieving Siri-based control typically involves a HomeKit bridge running on a Raspberry Pi — technically functional but outside the scope of what most households would configure independently.

Far-Field Microphone Performance in Singapore Apartments

The illuminated ring on Amazon Echo devices indicates wake word detection status — red indicates the microphone is muted, blue indicates active listening.

Singapore apartments present two acoustic challenges: hard tiled floors (standard in HDB and most condos) that create significant sound reflection, and AC unit noise. A typical wall-mounted split system running at medium fan speed generates approximately 38–42 dB of broadband noise at a distance of 2 metres.

The Echo Show 10's seven-microphone array handled wake word detection from 6 metres away in a tiled 20 sqm room with AC running — failure rate under 8%. The Echo Dot 4th Gen, tested in the same space, failed to detect the wake word at distances beyond 4 metres when the AC was running. The HomePod mini, positioned near a reflective corner, had the highest false activation rate — waking on television audio twice during a three-hour documentary session.

Privacy Considerations in Dense Housing

Always-on microphone devices in Singapore's dense residential environment raise a specific concern: wall and floor transmission of sound means conversations held at normal volume in one unit can sometimes be audible in adjacent units. None of the devices in this test stored locally processed audio — all rely on cloud processing with data centres outside Singapore. This conflicts with the data residency expectations of some users, particularly those with professional confidentiality concerns.

For users where this matters, local voice processing via Home Assistant with the Wyoming protocol and a Whisper speech-to-text model running locally represents the only fully private alternative currently available — at the cost of significantly lower recognition accuracy and requiring technical setup.

Summary Observations

• Google Assistant leads on Singapore English recognition and local data access, though gaps remain for NEA-specific queries.
• Alexa's smart home device ecosystem in Singapore is broader, but misrecognition behaviour (confident wrong responses vs. silent failures) warrants attention in automation contexts.
• Siri remains the weakest option for Singapore-specific use without significant HomeKit infrastructure investment.
• For AC control — the primary use case locally — all three work adequately when paired with the right hub or integration, but none work out of the box with the full range of Singapore-market units.
• Far-field microphone performance degrades materially with AC noise at distances beyond 4 metres for most sub-SGD 100 devices.