The latest release of ScanSoft Inc.s Dragon NaturallySpeaking voice recognition package provides nice incremental improvements but does not change the market significantly.
Version 7 of NaturallySpeaking, which began shipping at the end of March, provides users with more input options (including transcription of Pocket PC voice recordings) and more precise controls for improving accuracy and managing performance. The package provides an effective alternative data input and application command method for users who are more comfortable dictating than typing or for those who wish to avoid typing for health reasons.
NaturallySpeaking 7 runs on Windows NT 4.0, 98 SE, ME, 2000 and XP. NaturallySpeakings chief competitor is IBMs ViaVoice, but on April 25, ScanSoft announced it had reached a deal with IBM to take over worldwide sales, distribution and support for ViaVoice on the PC and Macintosh. IBM will continue to develop the product but will sell only versions of ViaVoice on nondesktop hardware.
eWEEK Labs has been using NaturallySpeaking 7 for the past two weeks. With practice in speaking clearly to get the best results and with the recognition accuracy settings set to their highest levels, we observed in our dictation tests an error rate of about one to two words for every 50 words (see screen). This is a decent accuracy rate and good enough for those who need or strongly desire voice recognition as a keyboard alternative.
However, the software—which ranges in price from $60 to about $200—is not yet at a place where casual users can rattle off a few quick e-mail messages without a good proofread. We may never get to this level of accuracy until software algorithms can advance significantly beyond the word-pair and word-triple probabilistic usage models on which current voice recognition software is based.
The $200 Solutions Series version we tested (also known as NaturallySpeaking Professional) is the only version with scripting and macro recorder features, digital recorder and Pocket PC support, and application command and control for Microsoft Corp.s Outlook.
This version also adds enterprise deployment features, including a network-installable .MSI (Microsoft Windows Installer) version and an administrator tool for centrally managing user profiles, custom vocabularies and voice command dictionaries (see screen).
Voice Variables
Voice Variables
Because so many variables affect speech recognition accuracy, its difficult to know for sure if the approach will work until users can try it for themselves in their own environments.
Officials at ScanSoft—which bought NaturallySpeaking from the bankrupt Lernout & Hauspie Speech Products N.V. in December 2001—claim a 15 percent improvement in accuracy in this release, due in part to an updated Vocabulary Optimizer feature that allows NaturallySpeaking to scan word processor documents or e-mail to determine typical word patterns (see screen).
Tuning NaturallySpeaking to a particular hardware environment was simpler with Version 7 than with past versions weve tested. Microphone testing and customization of recognition dictionaries took us only about 7 minutes.
The software includes six base language profiles to use as starting points for users that have American, British, Australian, Indian or Southeast Asian accents. Interestingly, the American accent profile came in adult and teen versions. (The teen version is customized to handle teen-agers higher-pitched voices.)
Misunderstanding
Misunderstanding
In testing, one of the biggest challenges was speaking so that we enunciated each word carefully enough.
We had initial problems with prepositions being omitted from text and with some words being consistently misinterpreted. (“Commoditization” and “may be” were problems for us.)
We also had trouble with verbs being recognized with the wrong tense (for example, “enable” versus “enabled”).
After adjusting our speaking patterns over time and making consistent use of the word recognition training facilities that come with NaturallySpeaking, we found that problem words were subsequently recognized correctly. This is a package that rewards persistence and user dedication.
We found that dictation worked best for creating initial drafts, but it wasnt well-suited for handling further document editing. There are good facilities for text selection and editing in the package, but moving text, joining and splitting sentences, and general navigation were faster with a keyboard than with equivalent voice commands.
An organizational issue that businesses need to weigh is whether the sound of dictation and correction will disturb nearby co-workers. Quiet, private offices are definitely the optimal environment for speech recognition.
Pocket PC Option
Pocket PC Option
NaturallySpeaking 7 is now certified for use with wireless headsets, array microphones (which can be placed on monitors or on desktops) and Pocket PC PDAs.
The package includes a voice recording application for Pocket PCs called ScanSoft Voice Recorder (see screen).
Voice recognition training on the Pocket PC was a lengthier process than on the PC: To train the system, we had to read a preselected passage for 15 minutes. The recording was then processed for about 25 minutes on our PC to generate a voice profile for Pocket PC recordings. Accuracy was noticeably worse when transcribing our Pocket PC-based recordings than when using a PC-based analog microphone, so we would not recommend using this option for anything other than first-draft transcriptions.
We had originally tried to use a third-party voice recording package we ordinarily use on the Pocket PC, but even though we recorded using the recommended format (pulse code modulation, 22.05KHz, 16-bit, mono), we couldnt see the file in NaturallySpeakings list of available audio files to recognize.
After installing ScanSoft Voice Recorder, we discovered that NaturallySpeaking is hard-coded to look only in a Pocket PC directory, “My Recordings,” which didnt exist on our machine until it was automatically created during the ScanSoft Voice Recorder install. Once we discovered this, we were able to use ScanSoft Voice Recorder and third-party software to create audio recordings for later transcription.
NaturallySpeaking 7 introduces a new feature where the software tries to automatically insert commas and periods into sentences rather than having users dictate their punctuation. In our tests, this function was right only about half the time, so after trying it for a few days, we abandoned the approach because it required too many corrections on our part.
West Coast Technical Director Timothy Dyck can be reached at timothy_dyck@ziffdavis.com.
Executive Summary
: Dragon NaturallySpeaking 7″>
Executive Summary: Dragon NaturallySpeaking 7 Solution Series
Usability |
Excellent |
Capability |
Good |
Performance |
Good |
Interoperability |
Fair |
Manageability |
Good |
Scalability |
Good |
Security |
Fair |
Organizations that want to provide dictation software to staff will find NaturallySpeaking 7 more effective and easier to manage than previous versions of the program. The $200 package had a good accuracy rate in eWEEK Labs tests and was faster at some tasks than its predecessor. Those wanting a cross-platform package should investigate IBMs ViaVoice. More information is available at www.scansoft.com/naturallyspeaking.
(+) Supports transcriptions from Pocket PC recordings and now supports array and wireless microphones; new enterprise deployment and management tools; improved accuracy; provides dictation capabilities and control of popular office applications using spoken commands.
(-) Automatic period- and comma-addition feature was hit or miss in tests; Pocket PC recording transcription recognition is less accurate than PC-based voice recognition; the package works only on Windows desktops. (A Pocket PC record-only client is also included.)
EVALUATION SHORT LIST
- IBMs ViaVoice
- Specialized voice recognition products for legal and medical use