Podcast by roles: how to create a dialogue with hosts and segments
The "Podcast by roles" tool creates a text dialogue for a podcast: you describe an idea, set host roles and desired duration, and receive a script with segments and natural transitions. This is not audio generation and not speech synthesis -- only text. In the article we will explain how to formulate queries, show educational examples and explain how to evaluate the result and what limitations to consider. All scripts in the article are fictional and provided as educational examples.
Open tool: Podcast script writer ↗How the tool works and what is input
The form consists of three fields: "Your text or idea" (up to 8000 characters), "Role or characters" (up to 1000 characters) and "Duration or duration". Additionally the output language is selected. After pressing the Generate button in a separate panel one result appears as plain text. It can be copied or downloaded as TXT. There are no style presets, sliders or levels of changes in the interface.
It is better to input a specific topic rather than a general "make a podcast". The more precisely you describe the subject, audience and desired structure, the more useful the draft will be. Host roles can be set as names and characters, e.g. "skeptic and enthusiast", or as expert positions: "practitioner and theorist". Duration can be indicated in minutes or words -- the tool orients on this parameter when distributing replicas.
Important to remember: the tool does not generate audio, video or images. It does not send your speech to AI -- if you use separate speech synthesis, that happens outside this service. Here you get only a script. Check that the topic does not require personal financial, medical or legal recommendations: such materials must remain educational.
Example query: input: "Podcast about how to start a home budget without stress", role: "Anya -- practitioner, Max -- skeptic", length: "5 minutes".
Scenario 1: educational podcast for beginners
The first example is a short episode for beginners. Task: explain a basic topic in simple words, bust a myth and give one practical step. Roles: host-enthusiast and co-host-skeptic. Duration: 5 minutes. In the query specify that three segments are needed: introduction, main part with example, conclusion with call to action.
Useful result should contain replicas that sound like live speech: short sentences, clarifying questions, light objections. Transitions between segments can be formatted as "Now let's review..." or "Let's return to the main point". If the answer has all replicas of the same length and no role switching, the request should be refined: add a requirement to alternate replicas and introduce no more than two ideas per segment.
Check: read aloud. If phrases are hard to pronounce or sound like an article rather than a conversation, ask to rewrite more conversationally. Make sure there are no fictional facts and statistics. The tool may make mistakes, so check all numbers and statements manually.
Result fragment: Anya: "Many think a budget is tables for half a day". Max: "Sounds boring. Is there a simpler way?" Anya: "Yes, start with one rule: record expenses for a week." Max: "And that's it?" Anya: "Just that. Then we add categories.".
Scenario 2: discussion podcast with two experts
The second example is a debate between experts on a professional topic. Task: show different points of view, but do not slide into conflict. Roles: "practitioner with experience in small business" and "analyst who looks at data". Duration: 10 minutes. In the query specify that four segments are needed: question setting, first argument, counter-argument, joint conclusion.
Useful result contains not only replicas, but also segment markers. For example: "Segment 1. What the question is". This helps with editing. Transitions should be natural: "You talk about experience, I propose to look at the numbers". If the tool gives too general phrases, add a specific industry and example situation to the request.
Check: make sure both voices are distinguishable by vocabulary and replica length. If both speak similarly, refine the characters: one uses metaphors, the other -- short facts. Remember, the script is a draft, not a ready air. It needs to be adapted to real hosts.
Example query: input: "Should a small business automate reporting", role: "Igor -- practitioner, Lena -- analyst", length: "10 minutes, 4 segments".
Scenario 3: interview with a guest
The third example is an interview, where one host asks questions and the other plays the guest role. Task: reveal the topic through personal experience. Roles: "host-journalist" and "guest -- teacher". Duration: 7 minutes. In the query specify that five blocks are needed: introduction, guest path, key lesson, mistake, advice to listeners.
Useful result should contain open-ended questions that require a detailed answer, not "yes/no". Transitions can be formatted as "Tell more..." or "What would you do differently?". If the guest's answers are too short, add a request to the query: each answer -- at least two sentences.
Check: read the questions aloud. They should sound respectful and not push to personal recommendations. If the topic concerns health, finances or law, make sure the material remains educational and does not give individual advice. The tool does not replace an editor and fact-checker.
Result fragment: Host: "Where did your interest start?" Guest: "From one book I took from the library." Host: "What turned out to be the most difficult?" Guest: "Not give up through a month".
How to evaluate the quality of the generated script
First criterion -- correspondence to the set roles. If you asked a skeptic to agree with everything, the result is weak. Second criterion -- structure: are there segments and transitions? Third -- naturalness of speech: short phrases, questions, pauses. Fourth -- absence of fictional facts and guarantees. Fifth -- compliance with duration: approximately 130-150 words per minute for conversational pace.
Check the text against the checklist: roles are distinguishable, segments are marked, transitions are logical, no repeats, no forbidden promises. If something is missing, do not rewrite everything manually -- refine the request and generate again. For example, add "avoid general phrases" or "add a specific example".
Remember the limitations: the model may make mistakes, queue and availability may change. The tool does not automatically save results, does not publish them and does not send them by email. Export -- only TXT. No DOCX, PDF, batch processing or integrations exist. This is important when planning work.
Checklist: 1) roles not mixed; 2) there are 3-5 segments; 3) transitions sound natural; 4) no fictional numbers; 5) length close to specified.
Typical errors in requests and how to fix them
First error: too general topic. "Podcast about life" gives a blurred result. Fix: narrow down to a specific question, e.g. "how to choose first profession". Second error: roles without characters. "Host 1 and Host 2" -- almost useless. Add adjectives: "cautious", "ironic", "pedantic".
Third error: duration not specified. Without it the tool can make too short or too long text. Indicate minutes or words. Fourth error: request to add statistics or quotes. The tool does not have access to sources and may invent data. Better to ask "leave places for facts I will check".
Fifth error: expecting audio. This tool creates only text. If you need sound, use separate speech synthesis, but remember: audio is not sent to AI within this service. Also do not expect photorealistic images -- the service works only with SVG and PNG vectors.
Bad request: "Make a podcast". Good request: "Make a 5-minute dialogue of two hosts about how to choose first profession; Anya -- practitioner, Max -- skeptic; three segments".
Tool limitations and ethical scenarios
The tool does not generate audio, video or photorealistic images. It does not clone voice, does not publish materials and does not send them by email. No registration, premium plan, batch processing or automatic saving. Maximum volume of first field -- 8000 characters, second -- 1000. For speech -- 3000 characters, for images -- 4 MB, for browser audio -- 20 MB or 5 minutes.
In topics about finances, health and law the script must remain educational. Do not ask the tool to give personal recommendations, set diagnoses or guarantee income. Formulate requests as "explain general principles" or "analyze typical errors". This reduces the risk of harmful advice.
All examples in the article are fictional. They show how a query and result may look, but are not real cases or research. Use them as a template for own experiments. Check every generated text before publication.
Ethical request: "Explain how compound interest works, without specific investment advice". Unethical: "Tell where to invest money to definitely earn"
The "Podcast by roles" tool helps quickly get a draft dialogue with roles, segments and transitions. Three examples -- educational episode, expert discussion and interview -- show how to formulate requests. Always check facts, adapt the text to real hosts and remember the limitations: text only, TXT export, no audio or photorealistic images. Scenarios from the article are fictional and provided for training.
Open tool: Podcast script writer ↗Interface translations are machine-assisted and may need refinement.