One of the most time-consuming parts of preparing a world language assessment is not always writing the questions. It is recording the audio.
A teacher must write a suitable script, find a quiet room, record it clearly, listen to the result, and perhaps record it again. A conversation activity requires several separate audio files. A dictation requires careful pacing. If a teacher needs another version for a make-up assessment, the whole process begins again.
There is also the problem of variety. Students who always hear their own teacher become accustomed to one voice, one accent, and one speaking rhythm. Authentic materials provide variety, but it can be difficult to locate a recording that matches the vocabulary, topic, length, and proficiency level of a particular lesson.
Innovation Assessments now includes AI assistance for generating world language audio activities inside the Test and Convo applications.
Teachers can create listening-comprehension passages, traditional dictations, and simulated conversations in fifteen languages:
- Arabic
- Chinese
- English
- French
- German
- Greek
- Hebrew
- Hindi
- Italian
- Japanese
- Korean
- Latin
- Portuguese
- Russian
- Spanish
The tools are not intended to remove the teacher from assessment design. They are designed to shorten the distance between an instructional idea and a usable classroom activity.
AI Listening Comprehension
The Test application now includes an AI Listening Comprehension generator.
The teacher selects:
- The target language.
- A CEFR proficiency level from A1 through C2.
- The student age or educational level.
- A short, medium, or long passage.
- A voice.
- A multiple-choice or short-answer question.
- A topic or scenario.
- Any additional instructions.
A teacher might request an A2 French passage about ordering breakfast in a café, a B1 Spanish announcement about a delayed train, or an A1 German description of a family.
The AI produces a draft containing both the listening script and a question based upon it. A multiple-choice item includes four answer choices and a designated correct response. A short-answer item includes model answers that may later assist with scoring.
The teacher sees this material before the audio is created.
This review step is important. AI can produce useful drafts, but it does not know precisely what a particular class has studied. It may use vocabulary that is too advanced, introduce a regional expression the teacher has not taught, or misunderstand an important detail in the request. The teacher can edit the script, question, choices, and answers before selecting Generate Audio + Add Question.
The finished audio and question are then added directly to the test.
When the student reaches the listening item, the passage plays automatically according to the Test application’s listening workflow. The student answers the multiple-choice or short-answer question without needing to leave the assessment.
Traditional Dictation
Dictation occupies an interesting place in language instruction. It is among the oldest methods used in the language classroom and, when used appropriately, it remains a useful exercise in listening discrimination, spelling, accents, punctuation, and the relationship between spoken and written language.
It is also tedious to record properly.
The AI Dictation generator uses a traditional three-pass structure. The completed recording is planned to:
- Read the passage naturally.
- Repeat it more deliberately, with clear separation and spoken punctuation.
- Read the complete passage naturally once more.
For several commonly taught languages, the system uses language-specific punctuation terms. A French dictation can say point, virgule, or point d’interrogation rather than inserting English punctuation instructions into the middle of the recording. Similar mappings are provided for Spanish, German, Italian, Portuguese, and English. Other supported languages use the general fallback behavior and therefore deserve especially careful teacher review.
The generated question asks students to write what they hear. The exact script becomes the primary model answer, while lightly normalized alternatives may also be included.
A teacher chooses the language, CEFR level, voice, passage length, age group, topic, and any special directions. As with listening comprehension, the complete draft can be edited before audio generation.
This makes it practical to create a short dictation aligned with the vocabulary of the current unit rather than searching for a preexisting recording that only approximately fits.
AI-Assisted Conversation Prompts
The Convo application presents a different problem.
Convo is designed to assess spontaneous speaking. The student hears one side of a conversation and records a response. The next prompt continues the situation until the student has completed a series of exchanges.
Preparing a good Convo requires more than writing several unrelated questions. The prompts must form a coherent interaction. They must provide enough context for a student to respond, but they must not supply the very content the student is expected to produce.
This last problem proved especially important during development.
An early AI-generated conversation about families in France included a prompt in which the conversation partner listed traditional, single-parent, and blended families. The corresponding student task was to name types of families. The audio had supplied the answer!
The generator has therefore been instructed to create only the conversation partner’s side. It must not state, preview, paraphrase, or provide examples of the student response being assessed.
The teacher describes the scenario and selects:
- One of the fifteen languages.
- A CEFR level from A1 through C2.
- A voice.
- Between two and eight prompts.
The AI creates a sequence of spoken turns forming one continuous conversation. These turns are not limited to direct questions. The conversation partner may make an observation, express a preference, offer an opinion, or describe a small problem that invites the student to react.
This makes the interaction sound more natural. Real conversations do not consist entirely of one person asking a list of interview questions.
Students Must Listen
Each generated Convo turn contains two different pieces of information.
The first is the actual spoken line the student hears. The second is a very brief visible cue such as:
- Respond naturally.
- React and explain.
- Answer and add one detail.
- Agree or disagree.
- Respond and ask a question.
The visible cue is intentionally vague. It should tell the student what kind of response to make without revealing the subject of the audio.
If the screen says, “Name three types of families in France,” the student does not need to understand the spoken French. The exercise has become a prepared speaking prompt rather than a listening-dependent conversation.
By using a cue such as “Answer with examples,” the student must comprehend the audio to know what examples are being requested.
This separation preserves the purpose of Convo: listening and responding in real time.
A Simulated Conversation, Not a Live AI Chat
The conversation is AI-generated, but it does not dynamically change according to what the student says.
The prompts are created in advance, reviewed by the teacher, converted to audio, and presented in a fixed sequence. The AI is not listening to the student and inventing the next turn during the assessment.
This is intentional.
A fixed sequence gives every student an equivalent task. It lets the teacher inspect the complete assessment beforehand. It also avoids the unpredictability, delay, and expense of running a live conversational AI during every student attempt.
The generator tries to maintain continuity without pretending to know what the student said. Later turns may use a content-free acknowledgment such as “I understand” or “That is interesting,” but they should not invent, summarize, or correct an unseen student response.
The result occupies a useful place between a disconnected list of speaking questions and a fully dynamic AI conversation.
Voice, Level, and Pacing
Both Test and Convo provide a selection of voice styles. The labels describe approximate personas—such as a calm adult voice, a warm male voice, or a polished female voice—rather than guaranteeing a particular regional identity.
Teachers can also adjust playback speed within a reasonable range. This can help match a recording to beginning or advanced learners without reducing speech to an unnatural crawl.
CEFR settings provide the AI with a useful target for vocabulary and sentence complexity. They should not be treated as an official certification that every generated sentence perfectly matches a proficiency level. As with all generated material, the teacher remains the final judge.
Pronunciation quality may also vary by language, name, regional expression, and selected voice. The fact that a language appears in the list means the system can be instructed to generate and speak it; it does not mean every voice will perform equally well in every language.
Listen before assigning!
Saving Preparation Time Without Surrendering Judgment
These tools perform several kinds of work at once. They can draft a passage, construct a question, produce answer choices or scoring models, and generate an audio file.
That is an impressive amount of assistance from a short teacher request.
It is also why review matters. An error in a private brainstorming response is inconvenient. An error converted into audio and placed on an assessment may confuse an entire class.
The workflow deliberately separates drafting from audio generation. Teachers can inspect and revise the material before using additional AI resources to create the recording. Generated questions and audio also become ordinary parts of the task afterward; the teacher can continue managing the assessment through the established Test or Convo tools.
AI usage is charged against the teacher’s available AI-token allowance. This gives subscribers control over how much generation they use and keeps the feature from silently producing unlimited external-service costs.
More Time for Designing the Assessment
The best use of artificial intelligence in education may not be to make instructional decisions for teachers. It may be to perform the mechanical work surrounding those decisions.
The teacher still decides what students should understand, which vocabulary belongs in the activity, how difficult the passage should be, what constitutes an acceptable answer, and whether the finished recording is appropriate.
The AI turns those decisions into a draft and a voice recording much faster than the traditional process.
For a language teacher who needs another listening passage, a carefully paced dictation, or five connected conversation prompts before tomorrow morning, that is no small improvement.
Authorship note: This feature announcement was generated by artificial intelligence using samples of David Jones’s blog writing as a stylistic guide. It was reviewed for consistency with the current Innovation Assessments Test and Convo AI-generation features.