Google Assistant

COMPANY

Google

ROLE

Creative Writer for NBU Market

Summary

In 3 months, I helped increase HI Google Assistant’s views by 50 million+ (277% increase).

 

PROBLEM STATEMENT

Hindi has an estimated 615 million global speakers (Ethnologue, 2011) and had 28 million views of Google Assistant Personality (January, 2019). We needed to create apt content and leverage this important NBU market.

CHALLENGES

  • Hindi’s 21+ dialects, intricately mixed in users’ language, was unrecognized by Google Assistant and resulted in losses.
  • HI’s half as many vowels and twice as many consonants, in comparison with English, led to ASR issues.
  • Arya’s mismatched clusters made coverage of losses difficult.
  • CMS Geppetto’s often archaic, grammatically incorrect and out-of-persona content corpus created a barrier in earning user trust.
  • GA’s wavenet voice’s robotic and often emotion-less prosody unfortunately got magnified when improving the prosody in Director version 1.

RESEARCH METHODS

Social Listening

Persona

NBU Content Strategy

Online Survey

Through close collaboration with the PgMs and leads, I evaluated the past two year’s social media updates on GA; studied the personality style guides, policies & principles; helped do an online survey of 30 Google dogfooders; and researched NBU content strategies.

The research showed there was a clear expectation for the HI Personality to be more a friend than an asistant. Light-hearted humor and empathetic writing was appreciated; and colloquialism prefered over formality.

DEFINING THE MVP

With the help of the cross-functional teams and the internal tools, Personality Dashboard, Geppetto & Director, I was able to accurately communicate subtle-but-necessary nuances in responses.

I etched out empathetic and witty responses right from the introductories such as, ‘Ok Google’, ‘How are you Google?’, setting the tone for a delightful conversation. Locale-specific Easter Eggs, their follow-up responses; and 60 basic, medium & discovery coverages of seasonal campaigns were created to establish trust and engagement.

STRATEGY

  • I added the mixed language losses as equivalent queries in Geppetto, covering all variations of user queries for a specific topic. 442,801 tags were labelled; 200,000+ losses added as EQs and new passages created.
  • Entire corpus (4000+ responses) was rewritten with a delightful and conversational language; 300+ out-of-persona, dated and biased responses sunset; and grammatical mistakes corrected.
  • Prosody of top 200 power passages was improved by modifying the transpose, tilt & variation in Director through self-voice with ‘copycat’ feature. Top 500 power passages were rewritten for voice, contextual words & punctuations were added, word syllables separated, silences & SFX added.
  • 4000+ ASR issues were tracked & labelled; elaborated on their semantic nature and broader speech team’s attention got on it. The issues are being investigated to improve ML.

data

RESULTS (April, 2019)

HI fetched 78.7 Million Views

(277% increase [50 M+])

77% Thumbs Up

(0.35% increase)

410 new responses received

10 Million Views, 76% Thumbs Up

data