Practice speeches, built around the interpreter.
Speech Library creates tailored, native-sounding speeches for simultaneous and consecutive interpreting—so practice can match the language, topic, terminology and challenge of the work ahead.
Get InterpretBankThe right speech is rarely the one you happen to find.
Practice speeches are indispensable in conference-interpreting programmes, yet pedagogically suitable material can be difficult to find and slow to write. Public collections—including the EU and UN speech repositories—are valuable resources, but cannot adapt to a learner’s language combination, current subject matter or readiness.
Speech Library gives students, teachers and professionals a straightforward way to create a new practice speech when they need one. Start with a topic; then set the conditions that make a speech useful for your particular training session.
Set the variables that shape interpreting difficulty.
Create speeches in multiple languages and combine the parameters below to move deliberately from a first exercise to a realistic conference scenario.
Language & topic
Choose the source language and the subject area you want to rehearse.
Difficulty & length
Adjust complexity and duration to fit the learning objective and available time.
Delivery speed
Work at a manageable pace, then raise the demand when you are ready.
Genre & register
Practise an address, briefing or other speaking situation in an appropriate voice.
Your glossary
Use terminology from the glossary you prepared for an event, in meaningful context.
Listen & review
Play or download the resulting speech for individual, classroom or peer practice.
Applications across the learning journey.
Students
Build repeated, graded exercises when suitable material is not readily available—and focus on the variables you are working to master.
Teachers
Prepare assignments that fit a learning objective, a class language combination and a chosen level of challenge.
Professionals
Rehearse an upcoming event with your own glossary terms, subject domain and a delivery style close to the assignment.
Putting AI practice speeches to the test.#
Creating a useful speech is more demanding than producing fluent text. We therefore carried out a formal, human-centred evaluation to examine whether Speech Library produces material that meets requested conditions, resembles expert-written training speeches and remains satisfying to use when delivered with an AI voice.
A formal evaluation grounded in expert practice.
How the study was designed.
For version 9.74, our team defined the construct of a pedagogically suitable practice speech from a corpus of 100 authentic speeches written by experts across 10 languages. We used those examples as in-context references to guide generation towards the features that make a speech interpretable—and productively challenging.
100 authentic, expert-written speeches in 10 languages.
Lexical and syntactic complexity, rhetorical devices and other features that influence interpreting difficulty.
A final blind evaluation with 50 users from student and professional interpreting populations.
Defining pedagogical suitability
The development study began with a question: how can a generated speech reflect the characteristics of material that experts write for interpreter training? Fluency alone is an incomplete answer. A useful practice speech also needs to correspond to the requested level of challenge and provide material appropriate to the exercise.
Our starting point was therefore human expertise. The 100 expert-written speeches provided the empirical reference for defining pedagogical suitability across 10 languages. This placed examples of actual training material at the centre of system design, with the aim of reproducing relevant characteristics in newly generated speeches.
From reference speeches to new material
The system uses in-context learning: reference examples guide the model during generation. The objective is to approximate the characteristics of expert-written practice speeches while producing content that responds to the user’s topic and parameters.
The design focuses on lexical complexity, syntactic complexity and rhetorical devices. These dimensions concern the vocabulary a speech uses, the structure of its sentences and the ways it develops an argument. They provide a more explicit basis for specifying difficulty than a general instruction to produce an “easy” or “advanced” text.
What we measured.
The final evaluation involved 50 InterpretBank users, including students and experienced interpreters. The blind comparison of human and generated speeches addressed perceived authorship, alongside assessment of whether generated material met the requested criteria. A further evaluation considered the speeches when delivered through AI voice synthesis.
1. Criterion adherence
Did generated material satisfy the requested characteristics? This outcome concerns the system’s control over the speech it produces. The reported result reached up to 96% adherence; it should be read as a maximum, rather than an overall success rate for all criteria and languages.
2. Perceived authorship
Could evaluators distinguish generated writing from human writing when its origin was concealed? Human speeches were sometimes attributed to AI, and generated speeches to humans. This task concerns perceived resemblance to human writing, rather than factual accuracy or demonstrated learning gains.
3. Voiced-speech experience
How did evaluators perceive the spoken output? Source identification and satisfaction on a five-point scale capture distinct outcomes: a voice may be recognisable as synthetic while still receiving a favourable satisfaction rating.
Strong satisfaction, with a clear limit on voice realism.
In the blind evaluation, generated written speeches followed the requested criteria in up to 96% of cases. Evaluators confused human and AI authorship in both directions, suggesting that the generated texts resembled expert-written practice material. The team reported no statistically significant difference between student and professional evaluators in their perceptions of the written speeches.
The voiced-speech result is different: 80% of testers correctly identified whether a speech was human or AI. The voice is therefore still recognisable as synthetic, yet 95% of satisfaction ratings fell in the top two categories. This distinction matters: voice realism and usefulness for rehearsal are related, but not the same outcome.
Taken together, the results support Speech Library as promising material for targeted practice alongside existing resources. They do not establish equivalent performance across every language, topic or proficiency level, or measure improvement in interpreting performance over time. Teachers and interpreters remain the final judges of whether a speech suits a particular exercise.
Evaluation results
Two complementary outcomes from the final user evaluation: how closely speeches followed the brief, and how listeners rated the AI-voiced delivery.
AI-voice satisfaction
Distribution of satisfaction ratings on a five-point scale.
Percentages reported from the final user evaluation (n = 50). “Up to 96%” reflects the highest reported accuracy across requested criteria. In the AI-voiced condition, 20% of satisfaction ratings were 5/5 and 75% were 4/5: together, 95% were in the two highest categories. The remaining ratings were distributed between 3/5 (3%) and 2/5 (2%). Separately, 80% correctly distinguished AI-voiced from human speech; this indicates that the voice remained identifiable as AI. Professionals were more critical than students in the voiced-speech assessment.
Make your next practice speech specific to you.
Speech Library is available in InterpretBank 9.74. Open a glossary, choose your parameters and create a practice session built around your work.
Download InterpretBank