google_speech_apis.md

Google Speech to Text API

Audio

Format:

• Format: FLAC
• Sample Rate: 16000Hz
• Chanels: 1 (mono)
• Depth: 16bit
• Bitrate: Unkwown limits, worked fine for 160Kbps

Converting to FLAC:

$ ffmpeg -i input_file.m4a input_file.mp3  # If needed (I recorded in a iPhone)
$ sox input_file.mp3 -r 16000 -b 16 -c 1 input_file.flac

API Call

Request

• Endpoint: https://www.google.com/speech-api/v1/recognize
• Method: POST
• Headers:
  • Content-Type: audio/x-flac; rate=16000;
  • User-Agent: Copy from a recent Chrome browser
• Query:
  • lang: pt-BR, en-US…
  • client: chrome
  • maxresults: integer
• Payload: Flac binary file.

Call:

$ curl -X POST --data-binary @audio_file.flac \
--header "Content-Type: audio/x-flac; rate=16000;" \
--user-agent "speech2text" \
"https://www.google.com/speech-api/v1/recognize?client=chromium&lang=pt-BR&maxresults=10"

Response

{
    "status": 0,
    "id": "ecdd8041e3eaee77c6897d0939389e40-1",
    "hypotheses": [
        {
            "utterance": "são paulo sim não celular wesley",
            "confidence": 0.77939147
        },
        {
            "utterance": "são paulo sim ou não celular wesley"
        }
    ]
}

Samples

Brazilian Portuguese

I asked some native brazilian portuguese speakers to say "São Paulo, sim, não, celular, [your name]", and got:

• "são paulo sim não celular wesley", confidence: 0.77939147
• "são paulo sim não celular rafael", confidence: 0.7373202
• "são paulo sim não celular roberta", confidence: 0.6814435
• "são paulo sim não celular pedro", confidence: 0.70905286
• "são paulo sim não celular diogo", confidence: 0.71847606

All testers were recorded using the same device (iPhone5) in similar noise conditions.

Spanish

Same test as above, but localized: "Santiago, si, no, celular, [your name]":

• "santiago si no celular fernanda", confidence: 0.6885493
  • Non-native speaker
  • Tested with Chile locale
• "Santiago si no celular Eduardo", confidence: 0.7460225
  • Non-native speaker
  • Tested as plain spanish without locale
  • Only case of proper names capilized

Hebrew

Same test as above, but localized: "Tel Aviv, yes, no, cellphone, [your name]":

• "תל אביב כן לא סלולרי שלומי אסף", confidence: 0.82569724
  • 100% accurate

Japanese

Same test as above, but localized: "Tokyo, yes, no, cellphone, [your name]":

• "東京 はい いいえ 携帯 ほな男", confidence: 0.4775479
  • Non-native speaker
  • Bad parse of non-japanese name
    • It didn't understand it was foreign
    • Should be parsed as "ロナウド" instead of "ほな男"
    • Probably fine for japanese names
  • All other words were fine
  • Confidence of 0.71819603 for the same audio without the name

Google Text to Speech API

API Call

Request

• Endpoint: https://translate.google.com/translate_tts
• Method: GET
• Headers:
  • User-Agent: Copy from a recent Chrome browser
• Query:
  • tl: pt-BR, en-US…
  • q: Texto to be spoken

Call:

$ curl -H 'User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_7_5) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/28.0.1500.71 Safari/537.36' \
'http://translate.google.com/translate_tts?tl=pt-BR&q=Rio%20de%20Janeiro' > output.mp3

Response

A MP3 file, with the following specification:

Format:

• Format: MP3
• Sample Rate: 22050Hz
• Chanels: 1 (mono)
• Depth: 16bit
• Bitrate: 32Kbps

Gender/Notes:

• Portuguese
  • pt-BR: female
  • pt-PT: female
• English
  • en-US: female
  • en-GB: male
• Spanish
  • es-MX: female
  • es-ES: female
  • es-CL: female
• Japanese
  • ja-JP: female (doesn't understand romaji, tries to spell every letter)
• Hebrew
  • he: no output