Blog Post

SpinVox Voicemail Transcription Done By Call Center Staff, Not Algorithm?

Stay on Top of Enterprise Technology Trends

Get updates impacting your industry from our GigaOm Research Community
Join the Community!

spinvoxlogoVoicemail transcription startup SpinVox is in a bit of a fix, after a BBC investigation found that the majority of messages the company claimed were converted to text by an advanced speech-to-text algorithm, were in fact “heard and transcribed by call center staff in South Africa and the Philippines.” SpinVox says that its technology “captures spoken words” and sends them to its proprietary algorithm to have them transcribed, but “when necessary, parts of messages can be sent to a ‘conversion expert.'” If true, it would be a huge blunder for the company to claim to be using an algorithm to convert calls, but really is using humans. It also calls into question the viability of speech-to-text conversions, including those at Google Voice, (s goog) which is notorious for having difficulties with its speech-to-text service.

The BBC found a Facebook group, allegedly created by staff at an Egyptian call center that once contracted with SpinVox, that included pictures of supposedly transcribed messages. The BBC also spoke with someone who claimed to work in one of those contracted call centers. That person claimed to have transcribed full messages, “from the very beginning to the very end” — which would go directly against what SpinVox has declared.

SpinVox vigorously denied the charges in a blog post today, saying that claims the company uses call centers to transcribe messages are “incorrect.” The company says that its service “know[s] what it doesn’t know” and can send only certain parts of a message to be inspected by humans — but that the messages are anonymized to prevent anyone from seeing the full message.

The company has had a rough month of July so far. It recently asked staff to be paid in stock rather than cash to try to save money — an especially curious move since SpinVox has raised more than $200 million from several companies including Goldman Sachs. The company said it would need more funding to finance growth from new deals it was making, including one with Telefonica in Latin America. It’s pretty rare for a news organization to directly call out a company like this without serious proof, especially one as prestigious as the BBC. It is even more rare for the company to flat-out deny the charges. I don’t know what to believe — but I think I’ll stick to old-fashioned listening-to-voicemails for now.

23 Responses to “SpinVox Voicemail Transcription Done By Call Center Staff, Not Algorithm?”

  1. kovakovsky

    hi, there’s also conversion center opened in Macedonia
    it is working a couple of months and it’s converting messages
    for SpinVox
    The contractor company name is Datalityx Europe
    Maybe 99 percent of the messages are converted by
    so called Quality Control Agent
    Currently messages are recived by UK and US

  2. Hi i want to know whether there are openings for VMT? i have worked as a VMT in a company called Focus MT, bangalore for a duration of 6 months. Can i apply, Please reply

  3. I want to know whether there are openings for VMT??
    I have worked in a company called Focus MT,bangalore as a VMT for 6 months. Can I get into this company ??
    Please reply.

  4. Hi… I’m Dennis and I’m one of the people from India who are hired by SpinVox (through a third party) to transcribe their voicemails to text. Certainly SpinVox does have an advanced speech-to-text algorithm system which they proudly boast of… but the messages which the computer or machine is unable to recognize/transcribe are then transferred to their centers/agents all around the world. These centers or agents are mainly situated in India, South Africa, Philippines and Egypt, and maybe in some other countries also. The tool which I or the agents all around the world use to transcribe SpinVox voicemails is called Tenzing.

  5. Actually, SpinVox has a software that captures voice and converts it to text or what we call voice recognition software. But of course, that software won’t be able to generate a 98% or greater, error-free transcripts so what they did was that they encorporated human transcribers with the system.
    How it works:
    1. Caller calls toll free number, leaves a message, then hangs up.
    2. Voicemail message is entered into the database, voice is captured by voice recognition software and is converted automatically into text.
    3. Converted text is transferred to human conversion agents for quality control (error checking like grammar, punctuation, sentence construction, etc).
    4. Voicemail message with the converted text message is transferred back to the database.

    I know because I worked as a conversion agent for Spinvox.
    Jott does it as well. Both are outsourcing here in the Philippines.

  6. spinvox is not the only one, i personally used to work for a company in the philippines where we transcribed SIMULSCRIBE’s voicemail messages into text!!!
    we were made to sign waivers and stuff for non-disclosure and things but the company treated us real bad and we got back by botching up the quality of the transcriptions.

  7. spinvox is not the only one, i personally used to work for a company in the philippines where we transcribed SIMULSCRIBE’s voicemail messages into text!!!

    we were made to sign waivers and stuff for non-disclosure and things but the company treated us real bad and we got back by botching up the quality of the transcriptions.

    subscribers have a right to know that their messages are being read by other people, in fact we were able to access almost everything, from the caller’s phone number to the person who “owns” the voicemails email and contact numbers.

    some people even sensitive information on the voicemails, pity the poor bugger who could get scammed because of this, i used to have co-workers who would copy the entire transcription including caller and called party contact details.

  8. This is a big duh. Just like Jott, Spinvox uses a combination of automated and human transcription. For anything complicated it’s going to go to a human. The key to making this profitable is to minimize the number of these that have to be done by a human, otherwise the labor costs eat you alive.

    Considering that Jott just exited to Nuance for what I would guess was around $10 mil., Spinvox has a huge whole to dig itself out of.

    • John Smith

      Spinvox is at least accurate in their transcripts, look at Nuance’s new VM2TXT service via Vonage. Despite Nuance’s claim that it has the best SR software on board they are getting all the Vonage voicemails transcribed at a facility located in Bangalore India by call center agents and their transcripts are also below par. If that is not infringement of data security for Vonage customers then what is? Nuance is delivering subpar quality with human transcriptionists is that fair?

  9. Not surprising in the least. The truly gullible folks are the investors who actually believed that this would work in a completely automated fashion, and therefore that there were breakaway economics of scale at play here. Voice Reco is really frickin’ difficult.

    Regardless of their level of disclosure to the public at a given moment, this probably disproves the fundamental business case for the company. With high customer acquisition costs, and high human transcription costs (clearly both are breaking the bank) is this sort of stand-alone business truly viable at its current price point?

    My guess is no. Great idea, though… but there are reasons why this was not implemented earlier.

  10. – if you are not concerned about the privacy of your voicemails and hate the fact that cellphone voicemail hasn’t really changed in 15 years then Spinvox is good for you. You must be completely gullible to believe that they had any kind of voice recognition since voicemail from an outdoor cellphone call has a match rate of 20% AT BEST.

    – the money issue is very very interesting (since it came from Goldman) which suggest that they were tied to a quick IPO and came with a bazillion of strings attached. my guess is that some were related to the growth rate of spinvox userbase, since they are not making the numbers they are probably trying to delay tapping into those funds to avoid dilution. quite sane on their part if it’s the case, I would do the same.

    I think there is a market for a product like spinvox but not at a 1/2 billion valuation.

  11. It depends on whether the people are English speaking natives or not, as to whether the description of their proprietary ‘algorithm’ has been misleading. See Searle’s Chinese room problem regarding Strong AI.

  12. WOW! Good post.

    1) If true, absolutely devastating. Wonder if there is any disclosure of human intervention in the T&C – i.e. Privacy?

    2) If true, wonder if the investors were aware of this. If not, I smell lawsuits coming.