#telephone — Public Fediverse posts on home.social

pezdessinateur @[email protected] · 2026-09-13 · 15:29 UTC

Le dessin du dimanche de La Dépêche du Midi sur les arnaques bancaires par téléphone #dimanche #arnaques #banque #argent #telephone #DessinDePresse #dessin

#dimanche #arnaques #banque #argent #telephone #dessindepresse

pezdessinateur @[email protected] · 2026-09-13 · 15:29 UTC

Le dessin du dimanche de La Dépêche du Midi sur les arnaques bancaires par téléphone #dimanche #arnaques #banque #argent #telephone #DessinDePresse #dessin

#dimanche #arnaques #banque #argent #telephone #dessindepresse

pezdessinateur @[email protected] · 2026-09-13 · 15:29 UTC

Le dessin du dimanche de La Dépêche du Midi sur les arnaques bancaires par téléphone #dimanche #arnaques #banque #argent #telephone #DessinDePresse #dessin

#dimanche #arnaques #banque #argent #telephone #dessindepresse

pezdessinateur @[email protected] · 2026-09-13 · 15:29 UTC

Le dessin du dimanche de La Dépêche du Midi sur les arnaques bancaires par téléphone #dimanche #arnaques #banque #argent #telephone #DessinDePresse #dessin

#dessin #dessindepresse #telephone #argent #banque #arnaques

Iran @[email protected] · 2026-09-13 · 09:51 UTC

https://www.europesays.com/iran/290086/ Iran, Turkey, Iraq Stress Need to Prevent Regional Escalation #Abbas #and #Araqchi #bilateral #Conflict #conversations #counterparts #Developments #diplomatic #discussing #efforts #Escalation #for #Foreign #further #held #his #Iran #iranian #Iraq #iraqi #Islam #Minister #need #of #Prevent #regional #RELATIONS #separate #stress #telephone #the #Times #to #Turkey #Turkish #with

#with #turkish #turkey #to #times #the

cob @[email protected] · 2026-09-12 · 14:20 UTC

🧵destinés aux 15-25 ans : #livre Pourquoi notre #cerveau adore scroller ? du Dr en neurosciences #ThibaudDumas

#France : moyenne du #temps passé sur #TikTok -> 1h23 /jour = 80 mètres de contenu déroulé par le pouce.

+ temps sur snap insta x fb yt etc.

#socialmedia #ia #addiction #psy #business #economie #year2026 #education #parents #enfants #telephone #sante #teameduc #science #emotion

#france #temps #tiktok #psy #business #economie

cob @[email protected] · 2026-09-12 · 14:20 UTC

🧵destinés aux 15-25 ans : #livre Pourquoi notre #cerveau adore scroller ? du Dr en neurosciences #ThibaudDumas

#France : moyenne du #temps passé sur #TikTok -> 1h23 /jour = 80 mètres de contenu déroulé par le pouce.

+ temps sur snap insta x fb yt etc.

#socialmedia #ia #addiction #psy #business #economie #year2026 #education #parents #enfants #telephone #sante #teameduc #science #emotion

#psy #business #economie #year2026 #education #parents

cob @[email protected] · 2026-09-12 · 14:20 UTC

🧵destinés aux 15-25 ans : #livre Pourquoi notre #cerveau adore scroller ? du Dr en neurosciences #ThibaudDumas

#France : moyenne du #temps passé sur #TikTok -> 1h23 /jour = 80 mètres de contenu déroulé par le pouce.

+ temps sur snap insta x fb yt etc.

#socialmedia #ia #addiction #psy #business #economie #year2026 #education #parents #enfants #telephone #sante #teameduc #science #emotion

#psy #business #economie #year2026 #education #parents

Nils Wilcke @[email protected] · 2026-09-07 · 07:26 UTC

Les deux tiers des collèges et lycées français ont une équipe incomplète pour cette rentrée, selon une enquête du Snes-FSU. Pourtant, le ministre de l'Education nationale, Edouard Geffray, avait affirmé quelques jours avant que les absences de profs seraient "marginales". Oui, mais au moins, les lycéens sont privés de téléphones... Les Macronistes ont érigé le mensonge politique en art

#Politique #Lycee #Macron #Emploi #EducationNationale #Rentree #Geffray #Gouvernement #Telephone #Tech

#politique #lycee #macron #emploi #educationnationale #rentree

Black Sail Studio @[email protected] · 2026-07-15 · 21:51 UTC

Building a Local AI Information-Gathering Phone Line

https://youtu.be/Yodq7gG_F3I

Modern voice agents are often presented as cloud services: a caller speaks, a remote platform transcribes the audio, a hosted language model generates a response, and another service synthesizes the reply.

For this project, I wanted to explore a different approach.

The goal was to build a locally controlled conversational phone-line prototype capable of gathering structured information from a caller while maintaining context throughout the interaction.

The system uses a locally hosted Gemma 4 model for dialogue, local speech recognition for transcription, and Kokoro 82M for text-to-speech. A custom Python application coordinates the entire interaction and presents it through a phone-style interface.

The result is an automated intake agent named Eva.

The Test Scenario

For the demonstration, Eva was configured to behave like a corporate information-gathering agent.

The caller began with a simple greeting. Eva then guided the conversation through a sequence of intake questions:

Full name
Company or organization
Callback number
Email address
Case, ticket, order, or reference number
General reason for the call
Product or service involved
Date and approximate time of the reported issue

The caller did not have to complete a rigid form or choose numbered menu options. Instead, the information was collected through natural conversation.

A portion of the interaction looked like this:

Eva: Hello. Thank you for contacting us. My name is Eva. To begin, may I please have your full name?
Caller: John Michael Smith.
Eva: Thank you, John Michael Smith. May I also have the company or organization you are calling from?

As the conversation continued, Eva retained the previously supplied information and advanced to the next relevant question.

When the caller explained that they were upset about new automated systems, Eva did not restart the intake process or lose track of the conversation. She asked a follow-up question to determine which product or service was involved.

That contextual continuity is the most important part of the demonstration.

The Core Architecture

The system is built as a real-time conversational pipeline:

 Microphone input        ↓  Speech-to-text        ↓  Conversation controller        ↓  Gemma 4 through LM Studio        ↓  Kokoro 82M text-to-speech        ↓  Audio playback

Each component has a distinct responsibility.

The speech-to-text layer converts the caller’s voice into text. The conversation controller maintains the active call state and sends the transcript to the language model. Gemma 4 evaluates the current message alongside the prior conversation and produces the next response. Kokoro 82M then converts that response into spoken audio.

The Python application ties these systems together.

It manages:

Microphone capture
Voice activity detection
Turn-taking
Conversation history
System prompts
Model requests
Voice generation
Audio playback
Call state
Session termination
Persistent memory
Transcript logging

The language model is only one component of the system. The surrounding orchestration is what makes the experience feel like a coherent call rather than a sequence of unrelated AI requests.

Why Conversation History Matters

A language model does not automatically remember everything that happened earlier in a call.

The application must preserve the conversation and provide the relevant history with each new request.

Without that context, an information-gathering agent might repeatedly ask for the caller’s name, forget which fields were already completed, or ask questions in an inconsistent order.

In this prototype, the conversation controller tracks both sides of the exchange:

 System instructions  Caller message  Agent response  Caller message  Agent response

That history allows the model to understand which information has already been collected and what still needs to be asked.

It also enables more natural follow-up behavior.

For example, when the caller says:

“I am upset about the new automated systems.”

Eva can interpret that statement as the reason for the call and ask which product is affected, rather than simply continuing through an unrelated checklist.

Structured Behavior Without a Traditional Form

One of the interesting aspects of this design is that the model can follow a structured intake objective without forcing the caller through a conventional form.

The system prompt defines the agent’s responsibilities, tone, required information, and conversational boundaries.

The agent can then gather the same fields a form would collect while allowing the caller to speak naturally.

This creates a hybrid between two familiar systems:

 Rigid automated phone menu              +  Human-style conversational intake

The structured requirements remain in place, but the interface becomes conversational.

That can be useful in situations where callers may not know how their problem should be categorized before they begin speaking.

Local Inference and Control

The language model is served locally through LM Studio rather than being accessed through a commercial cloud API.

This provides direct control over:

Model selection
System prompts
Conversation history
Context length
Generation settings
Data retention
Personality configuration
Memory behavior

The text-to-speech component is also hosted locally.

That means the central conversation does not depend on sending every transcript and model response to an external AI provider.

Local inference introduces its own technical requirements, including GPU resources, model management, latency optimization, and service coordination. However, it also gives the developer considerably more control over the complete conversational stack.

Voice Activity Detection

The caller does not need to press a push-to-talk button during normal operation.

The application monitors the microphone and determines when speech begins and ends.

This process is known as voice activity detection.

The system moves through several internal states:

 Listening     ↓  Speech detected     ↓  Recording     ↓  Silence detected     ↓  Transcribing     ↓  Thinking     ↓  Speaking     ↓  Listening

This state-based design is important because the microphone, speech model, language model, and voice model all operate at different speeds.

The application must know which component currently owns the interaction.

It must also prevent the microphone from transcribing the AI’s own voice as new caller input.

Persistent Memory

The application also includes persistent conversational memory.

This allows the agent to retain selected context between separate sessions rather than beginning every call as a completely blank system.

The memory architecture separates several types of information:

Current-call conversation
Recent message history
Summaries of earlier calls
Long-term memories
Archived transcripts

This distinction matters because sending every previous conversation back into the model would eventually become inefficient.

Instead, recent dialogue can remain verbatim while older interactions are summarized. Important details can be stored separately and included only when relevant.

The application, not the language model, owns this memory.

That is an important architectural principle.

The model generates language, but the surrounding software determines what is saved, retrieved, discarded, or presented as context.

The Interface

The application was designed to resemble a live phone call rather than a conventional chatbot window.

The interface includes:

Start and hang-up controls
Call timer
Current system state
Editable system prompt
Agent profiles
Live transcription
Model responses
Diagnostic timing information
Persistent memory controls
A reactive circular audio visualization

The visualizer changes according to the active state.

The caller’s speech is represented in yellow. Model processing uses blue and purple states. The generated AI voice is shown in green.

This is primarily an aesthetic feature, but it also provides immediate feedback about what the system is doing.

A user can see whether the application is listening, transcribing, waiting for the model, generating speech, or playing audio.

What This Demonstration Proves

This experiment demonstrates that a locally hosted language model can operate as the conversational core of an automated information-gathering line.

More specifically, it shows that the system can:

Listen for natural speech
Convert speech into text
Preserve conversational context
Collect information in sequence
Ask relevant follow-up questions
Produce spoken responses
Maintain a consistent agent identity
Store and recall previous conversational information
Operate through a custom local interface

The prototype is not intended to replace a production call center in its current form.

A production deployment would require additional work in areas such as authentication, encryption, regulatory compliance, consent, auditing, data validation, failure recovery, telephony integration, and human escalation.

But the central conversational mechanism is functional.

The Larger Possibility

The broader significance of this project is not limited to corporate intake.

The same architecture could support many types of locally controlled voice agents:

Technical support intake
Appointment pre-screening
Internal help desks
Interactive training systems
Voice-based surveys
Game characters
Accessibility tools
Personal assistants
Research interviews
Educational tutors

The underlying pattern remains the same:

 Listen  Understand  Maintain context  Respond  Remember

What changes is the system prompt, the information being collected, the voice, and the surrounding workflow.

Final Thoughts

The project began as a simple idea: speak into a microphone, send the transcript to a local model, and play the response through a local voice engine.

Once persistent memory, automatic voice detection, agent profiles, structured prompts, call archives, and state management were added, it became something more substantial.

It became a locally controlled conversational-agent platform.

The most important lesson from the project is that the model itself is not the complete system.

A useful voice agent emerges from the coordination of perception, context, reasoning, memory, expression, and interface design.

Gemma 4 provides the conversational intelligence.

Kokoro 82M provides the voice.

The Python application provides the structure that allows them to behave like a single coherent agent.

-Me 7-15-2026

#agent #ai #artificialIntelligence #assistant #call #llm #operator #phone #phonecall #technology #telephone

#agent #ai #artificialintelligence #assistant #call #llm

Arbrealettres @[email protected] · 2026-07-10 · 14:16 UTC

La guerre a pris fin (Nasser Rabbah)

Illustration La guerre a pris fin La guerre a pris fin J'ai examiné mon corps ma tête, mes doigts et mes bras Rien ne manquait comme si tout avait regagné sa place sur-le-champ J'ai regardé le ciel C'est quoi, là ? De la fumée ou des nuages ? Les oiseaux me manquent pas les avions La guerre a pris fin Le balai a enlevé la poussière, les bris de verre, les clous de la porte cassée, les cailloux réduits en gravats, les éclats des verres de thé, le cadre de […]

https://arbrealettres.wordpress.com/2026/07/10/la-guerre-a-pris-fin-nasser-rabbah/

#nasserrabbah #accueillir #ami #arriver #avion #eclat

MediaFaro Magazine @[email protected] · 2026-07-01 · 10:40 UTC

Finland’s last analogue landline phones go silent after 150 years.

Finland’s last landline call has been made as the Nordic country becomes the latest to retire copper-wire phones in favour of fibre.

Estonia, the Netherlands, Norway and Spain have already made the jump as countries across the world roll out fibre optic cable that can handle both internet services and voice calls.

https://mediafaro.org/article/20260630-finlands-last-analogue-landline-phones-go-silent-after-150-years?mf_channel=mastodon&action=forward

#Finland #Landline #Telephone #FibreOptic #POTS #Analogue