The problem
People’s Law School wanted to solve two problems. First, their Beagle+ chatbot answered questions about the law in BC, but only if people typed them. Typing out a long legal question could be slow and inconvenient.
Second, Dial-A-Law, People’s Law School’s sister site, read pages aloud with a computer-sounding voice, but newer AI voices sound far more natural. The recordings on the Dial-A-Law phone line used an older voice too.
The challenge
The obvious fix for the chatbot was real-time voice, where a person and an AI talk back and forth like a phone call. But real-time voice can be unforgiving and expensive. People expect a reply within a few hundred milliseconds, even while the AI searches a legal database. The AI also can’t cut in when someone pauses mid-sentence to think. Tangowork estimated real-time voice could cost People’s Law School up to $800 a month, too much for a non-profit.
A spoken question in Beagle+, from recording to the answer being read aloud. The transcript appears in the chat as the person’s question, so they can see what Beagle+ heard, and it’s answered right away. (We asked this question ourselves for the screenshot.)
The solution
Tangowork chose a simpler design at a fraction of the cost: record, transcribe, answer, then read aloud on request. OpenAI gpt-4o-mini-transcribe turns a spoken question into text. From there, the question is handled exactly like a typed one, with the same search, prompts, and guardrails. A speaker button on every answer reads it aloud with gpt-4o-mini-tts, in English, French, or Mandarin.
On Dial-A-Law, the same AI voice replaced the old read-aloud service. When an editor publishes a page in Contentful, the translation pipeline we built earlier updates the French and Simplified Chinese versions. New audio in all 3 languages usually follows within 4 minutes. Each recording is labelled with a fingerprint of its words, so a page is re-recorded only when its words change. The Dial-A-Law phone line now plays the same recordings.
Every Dial-A-Law topic page has a “Listen to this page” player in English, French, and Chinese, with a button to download the MP3.
English
Buying a homeFrench
Acheter une maisonMandarin
购买住房
The first 55 seconds of the Dial-A-Law guide “Buying a home”, narrated by AI in each language. The Simplified Chinese page is read in Mandarin. The full recordings run 7 to 9 minutes.
The results
- About 2 cents a conversation. Transcribing questions and reading answers aloud adds about 2 cents to the cost of a conversation. Beagle+ offers 100 chats a day. If every single one used voice, that would add about $50 a month, compared with up to $800 for real-time voice.
- 96 hours of narration for about $87. All 215 Dial-A-Law topic pages have recordings in English, French, and Mandarin. That’s about 40 cents a page for all 3 languages.
- Answers start playing in about 1.3 seconds. In most browsers, the audio streams as it’s generated, so there’s no wait for the whole file. Transcribing a spoken question takes about as long.
- People’s Law School keeps no voice recordings. A spoken question is turned into text, and the audio is discarded. The transcript is kept, and software masks personal information in it, so spoken questions get the same accuracy reviews as typed ones.