I-Vibe Ikhodelwe Ithuluzi Elihlaziya Imizwa Yekhasimende Nezihloko Ezivela Ekurekhodweni Kwezingcingo

# Isingeniso

Nsuku zonke, izikhungo zesevisi yamakhasimende zirekhoda izinkulungwane zezingxoxo. Kufihlwe kulawo mafayela alalelwayo kunemigodi yegolide yolwazi. Ingabe amakhasimende anelisekile? Yiziphi izinkinga abavame ukuzikhuluma? Imizwa ishintsha kanjani phakathi nocingo?
Ukuhlaziya lokhu okurekhodiwe mathupha kuyinselele. Nokho, ngobuhlakani besimanje bokwenziwa (AI), singabhala amakholi ngokuzenzekelayo, sithole imizwa, futhi sikhiphe izihloko ezivelayo — konke okungaxhunyiwe ku-inthanethi nangamathuluzi omthombo ovulekile.

Kulesi sihloko, ngizokuhambisa ngephrojekthi ephelele yokuhlaziya imizwa yekhasimende. Uzofunda ukuthi:

• Ukuloba amafayela omsindo kumbhalo usebenzisa Hleba
• Ukuthola imizwa (emihle, engemihle, engathathi hlangothi) kanye nemizwa (ukukhungatheka, ukwaneliseka, ukuphuthuma)
• Ikhipha izihloko ngokuzenzakalelayo BERIsihloko
• Ibonisa imiphumela kudeshibhodi esebenzisanayo

Ingxenye engcono kakhulu ukuthi yonke into isebenza endaweni. Idatha yakho ebucayi yekhasimende ayilokothi ishiye umshini wakho.

I-Fig 1: Uhlolojikelele lwedeshibhodi ebonisa igeji yemizwa, i-radar yemizwa, nokusabalalisa kwesihloko

# Ukuqonda Kungani I-AI Yasendaweni Ibalulekile Kudatha Yekhasimende

Izinsizakalo ze-AI ezisekelwe efwini ezifana I-OpenAI's API zinamandla, kodwa ziza nokukhathazeka okufana nezindaba zobumfihlo, lapho izingcingo zamakhasimende ngokuvamile ziqukethe ulwazi lomuntu siqu; izindleko eziphezulu, lapho ukhokha khona intengo yekholi ngayinye ye-API, engeza ngokushesha ngamavolumu aphezulu; kanye nokuncika emikhawulweni yezinga le-inthanethi. Ngokusebenza endaweni, kulula ukuhlangabezana nezidingo zokuhlala kwedatha.

Lesi sifundo se-AI senkulumo-kuya-umbhalo sasendaweni sigcina yonke into ku-hardware yakho. Amamodeli alanda kanye futhi asebenza engaxhunyiwe ku-inthanethi unomphela.

Umfanekiso wesi-2: Uhlolojikelele lwe-Architecture yesistimu ebonisa ukuthi ingxenye ngayinye iwuphatha kahle umsebenzi owodwa. Lo mklamo we-modular wenza isistimu iqondeke kalula, ihlolwe, futhi inwetshwe

// Okudingekayo

Ngaphambi kokuqala, qiniseka ukuthi unokulandelayo:

• I-Python 3.9+ ifakwe emshinini wakho.
• Kufanele ube nakho FFmpeg efakelwe ukucutshungulwa komsindo.
• Kufanele ube nokujwayela okuyisisekelo ngePython nemiqondo yokufunda komshini.
• Udinga cishe u-2GB wesikhala sediski samamodeli e-AI.

// Ukusetha Iphrojekthi Yakho

Vala indawo yokugcina futhi usethe indawo yakho:

git clone 

Dala indawo ebonakalayo:

Yenza kusebenze (iWindows):

Yenza kusebenze (Mac/Linux):

Faka okuncikile:

pip install -r requirements.txt

I-run yokuqala ilanda amamodeli we-AI (~1.5GB isamba). Ngemva kwalokho, yonke into isebenza ungaxhunyiwe ku-inthanethi.

Umfanekiso wesi-3: Itheminali ekhombisa ukufakwa ngempumelelo

# Iloba umsindo nge-Whisper

Kumhlaziyi wemizwa yekhasimende, isinyathelo sokuqala ukuguqula amagama akhulunywayo asuke ekurekhodweni kwekholi abe umbhalo. Lokhu kwenziwa yi-Whisper, isistimu yokuqaphela inkulumo ezenzakalelayo (ASR) eyakhiwe ngu I-OpenAI. Ake sibheke ukuthi isebenza kanjani, kungani kuyisinqumo esihle, nokuthi siyisebenzisa kanjani kuphrojekthi.

I-Whisper imodeli ye-encoder-based encoder-decoder eqeqeshwe amahora angu-680,000 omsindo wezilimi eziningi. Uma uyiphakela ifayela lomsindo, ithi:

• Izama kabusha umsindo ube yi-16kHz mono
• Ikhiqiza i-mel spectrogram – ukumelwa okubonakalayo kwamafrikhwensi ngokuhamba kwesikhathi – esebenza njengesithombe somsindo
• Ihlukanisa i-spectrogram ibe amawindi amasekhondi angu-30
• Idlula iwindi ngalinye kusifaki khodi esidala izethulo ezifihliwe
• Kuhunyushwa lezi izethulo zibe amathokheni ombhalo, igama elilodwa (noma igama elincane) ngesikhathi

Cabanga nge-mel spectrogram njengoba imishini “ibona” ​​izwakala kanjani. I-x-eksisi imele isikhathi, i-axis ka-y imele imvamisa, futhi ukushuba kombala kubonisa ivolomu. Umphumela uba umbhalo onembe kakhulu, ngisho nangomsindo wangemuva noma iziphimiso.

Ukusebenzisa Ikhodi

Nansi indlela yokubhala ewumongo:

import whisper

class AudioTranscriber:
    def __init__(self, model_size="base"):
        self.model = whisper.load_model(model_size)
   
    def transcribe_audio(self, audio_path):
        result = self.model.transcribe(
            str(audio_path),
            word_timestamps=True,
            condition_on_previous_text=True
        )
        return {
            "text": result["text"],
            "segments": result["segments"],
            "language": result["language"]
        }

I model_size ipharamitha ilawula ukunemba ngokumelene nesivinini.

Ezimweni eziningi zokusetshenziswa, base noma small inikeza ibhalansi engcono kakhulu.

Umdwebo 4: Okuphumayo kokulotshiweyo okubonisa amasegimenti anesitembu sesikhathi

# Ukuhlaziya Imizwa Ngeziguquli

Ngombhalo okhishiwe, sihlaziya imizwa sisebenzisa Ama-Face Transformers Okwanga. Sisebenzisa iCardiffNLP's RoBERTa imodeli, oqeqeshwe kumbhalo wenkundla yezokuxhumana, olungele izingcingo zekhasimende ezixoxayo.

// Ukuqhathanisa Imizwa Nemizwa

Ukuhlaziya imizwa kuhlukanisa umbhalo njengophozithivu, ongathathi hlangothi, noma onegethivu. Sisebenzisa imodeli ye-RoBERTa ecushwe kahle ngoba iqonda kangcono umongo kunokufanisa amagama angukhiye alula.

Okulotshiweyo kwenziwa ithokheni futhi kudluliswe ku-Transformer. Isendlalelo sokugcina sisebenzisa ukwenza kusebenze i-softmax, ekhipha amathuba afinyelela ku-1. Isibonelo, uma phozithivu ingu-0.85, i-negetive ingu-0.10, futhi inegethivu ingu-0.05, lapho-ke usuwonke umuzwa uthi positive.

• Imizwa: I-polarity isiyonke (enhle, embi, noma engathathi hlangothi) ephendula umbuzo: “Ingabe lokhu kuhle noma kubi?”
• Umzwelo: Imizwa ethize (intukuthelo, injabulo, ukwesaba) ephendula umbuzo: “Yini ngempela abayizwayo?”

Sithola kokubili ukuze uthole ukuqonda okuphelele.

// Ukusebenzisa Ikhodi Yokuhlaziya Imizwa

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch.nn.functional as F

class SentimentAnalyzer:
    def __init__(self):
        model_name = "cardiffnlp/twitter-roberta-base-sentiment-latest"
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.model = AutoModelForSequenceClassification.from_pretrained(model_name)
   
    def analyze(self, text):
        inputs = self.tokenizer(text, return_tensors="pt", truncation=True)
        outputs = self.model(**inputs)
        probabilities = F.softmax(outputs.logits, dim=1)
       
        labels = ["negative", "neutral", "positive"]
        scores = {label: float(prob) for label, prob in zip(labels, probabilities[0])}
       
        return {
            "label": max(scores, key=scores.get),
            "scores": scores,
            "compound": scores["positive"] - scores["negative"]
        }

I compound amaphuzu asukela ku -1 (okubi kakhulu) ukuya +1 (okuhle kakhulu), okwenza kube lula ukulandelela izitayela zemizwa ngokuhamba kwesikhathi.

// Kungani Ugwema Izindlela Ezilula Ze-Lexicon?

Izindlela zendabuko ezifana I-VADER bala amagama akhayo naphikisayo. Nokho, bavame ukugeja umongo:

• “Ayiyinhle lento.” I-Lexicon ibona “okuhle” njengokuhle.
• I-transformer iqonda ukunganaki (“hhayi”) njengokunegethivu.

Ama-Transformer aqonda ubudlelwano phakathi kwamagama, okuwenza anembe kakhulu kumbhalo womhlaba wangempela.

# Ikhipha Izihloko nge-BERTopic

Ukwazi imizwa kuyasiza, kodwa amakhasimende akhuluma ngani? BERIsihloko ithola ngokuzenzakalela amatimu embhalweni ngaphandle kokuthi uwachaze kusengaphambili.

// Isebenza kanjani i-BERTopic

• Ukushumeka: Guqula umbhalo ngamunye ube ivekhtha usebenzisa Umusho Transformers
• Ukuncishiswa kobukhulu: UMAP Icindezela lawa ma-vector endaweni enohlangothi oluphansi
• Ukuhlanganisa: I-HDBSCAN ihlanganisa imibhalo efanayo ndawonye
• Ukumelwa Kwesihloko: Kuqoqo ngalinye, khipha amagama afaneleka kakhulu usebenzisa i-c-TF-IDF

Umphumela uba isethi yezihloko ezifana “nezinkinga zenkokhelo,” “usekelo lwezobuchwepheshe,” noma “impendulo yomkhiqizo.” Ngokungafani izindlela ezindala like I-Latent Dirichlet Allocation (LDA)BERIsihloko siqonda incazelo ye-semantic. “Ukubambezeleka komkhumbi” kanye “nokulethwa sekwephuzile” kuhlangana ndawonye ngoba anencazelo efanayo.

Ukusebenzisa Ikhodi

Kusuka topics.py:

from bertopic import BERTopic

class TopicExtractor:
    def __init__(self):
        self.model = BERTopic(
            embedding_model="all-MiniLM-L6-v2",
            min_topic_size=2,
            verbose=True
        )
   
    def extract_topics(self, documents):
        topics, probabilities = self.model.fit_transform(documents)
       
        topic_info = self.model.get_topic_info()
        topic_keywords = {
            topic_id: self.model.get_topic(topic_id)[:5]
            for topic_id in set(topics) if topic_id != -1
        }
       
        return {
            "assignments": topics,
            "keywords": topic_keywords,
            "distribution": topic_info
        }

Qaphela: Ukukhishwa kwesihloko kudinga imibhalo eminingi (okungenani 5-10) ukuze kutholwe amaphethini anengqondo. Izingcingo ezingazodwa zihlaziywa kusetshenziswa imodeli efakiwe.

Umdwebo 5: Ishadi lebha yokusabalalisa isihloko elibonisa izigaba zokukhokhisa, ukuthunyelwa, kanye nosekelo lobuchwepheshe

# Ukwakha Ideshibhodi Esebenzisanayo nge-Streamlit

Idatha eluhlaza kunzima ukuyicubungula. Sakha a Sakaza ideshibhodi (app.py) evumela abasebenzisi bebhizinisi ukuthi bahlole imiphumela. I-Streamlit ishintsha imibhalo yePython ibe izinhlelo zokusebenza zewebhu ezinekhodi encane. Ideshibhodi yethu inikeza:

• Layisha isixhumi esibonakalayo samafayela alalelwayo
• Ukucubungula kwesikhathi sangempela ngezinkomba zenqubekelaphambili
• Ukuboniswa okusebenzisanayo kusetshenziswa Icebo
• Ikhono lokuya phansi lokuhlola amakholi angawodwana

// Ukusetshenziswa Kwekhodi Yesakhiwo Sedeshibhodi

import streamlit as st

def main():
    st.title("Customer Sentiment Analyzer")
   
    uploaded_files = st.file_uploader(
        "Upload Audio Files",
        type=["mp3", "wav"],
        accept_multiple_files=True
    )
   
    if uploaded_files and st.button("Analyze"):
        with st.spinner("Processing..."):
            results = pipeline.process_batch(uploaded_files)
       
        # Display results
        col1, col2 = st.columns(2)
        with col1:
            st.plotly_chart(create_sentiment_gauge(results))
        with col2:
            st.plotly_chart(create_emotion_radar(results))

Ukulondoloza isikhashana kwe-Streamlit @st.cache_resource iqinisekisa ukuthi amamodeli alayisha kanye futhi aqhubeke kukho konke ukusebenzelana, okubalulekile kulwazi lomsebenzisi oluphendulayo.

Umdwebo 7: Ideshibhodi egcwele enezinketho zebha eseceleni namathebhu amaningi okubuka

// Izici Eziyinhloko

• Layisha umsindo (noma sebenzisa imibhalo eyisampula ukuhlola)
• Buka okulotshiweyo okunokugqamisa kwemizwelo
• Umugqa wesikhathi womzwelo (uma ucingo lude ngokwanele)
• Ukubukwa kwesihloko kusetshenziswa amashadi asebenzisanayo e-Plotly

// Ukulondoloza isikhashana kokusebenza

I-Streamlit iqalisa kabusha iskripthi kukho konke ukuhlanganyela. Ukugwema ukucubungula kabusha amamodeli asindayo, sisebenzisa @st.cache_resource:

@st.cache_resource
def load_models():
    return CallProcessor()

processor = load_models()

// Ukucubungula Kwesikhathi Sangempela

Uma umsebenzisi elayisha ifayela, sibonisa isipina ngenkathi sicubungula, bese siveza imiphumela ngokushesha:

if uploaded_file:
    with st.spinner("Transcribing and analyzing..."):
        result = processor.process_file(uploaded_file)
    st.success("Done!")
    st.write(result["text"])
    st.metric("Sentiment", result["sentiment"]["label"])

# Ukubukeza Izifundo Ezisebenzayo

Ukucutshungulwa komsindo: Ukusuka ku-Waveform kuya kumbhalo

Umlingo we-Whisper usenguqukweni yawo ye-mel spectrogram. Ukuzwa komuntu kuyi-logarithmic, okusho ukuthi singcono kakhulu ekuboneni amaza aphansi kunalawo aphezulu. Isikali se-mel silingisa lokhu, ngakho imodeli “izwa” kakhulu njengomuntu. I-spectrogram iyisithombe se-2D (isikhathi siqhathaniswa nemvamisa), isishumeki se-Transformer esisicubungula ngendlela efanayo nendlela esizocubungula ngayo isiqeshana sesithombe. Yingakho i-Whisper iwuphatha kahle umsindo onomsindo; ibona isithombe sonke.

// I-Transformer Output: I-Softmax vs. Sigmoid

• I-Softmax (imizwa): Iphoqelela amathuba okuba enze isamba esingu-1. Lokhu kulungele izigaba ezihlukene, njengoba umusho ngokuvamile ungekho kokubili ophozithivu noma onegethivu.
• I-Sigmoid (imizwa): Uphatha ikilasi ngalinye ngokuzimela. Umusho ungajabulisa futhi umangale ngesikhathi esifanayo. I-Sigmoid ivumela lokhu kugqagqana.

Ukukhetha ukwenza kusebenze okulungile kubalulekile esizindeni sakho esinenkinga.

// Ukuxhumana Ngemininingwane Ngokubona

Ideshibhodi enhle yenza okungaphezu kokubonisa izinombolo; ixoxa indaba. Amashadi esakhiwo ayasebenzisana; abasebenzisi bangahambisa phezulu ukuze babone imininingwane, basondeze kububanzi besikhathi, futhi bachofoze izinganekwane ukuze uguqule uchungechunge lwedatha. Lokhu kuguqula izibalo ezingahluziwe zibe imininingwane esebenzisekayo.

// Ukuqalisa Isicelo

Ukuze usebenzise uhlelo, landela izinyathelo ezisuka ekuqaleni kwalesi sihloko. Hlola ukuhlaziya imizwelo ngaphandle kwamafayela alalelwayo:

Lokhu kusebenzisa isampula yombhalo ngamamodeli we-natural language processing (NLP) futhi kubonisa imiphumela kutheminali.

Hlaziya okurekhodiwe okukodwa:

python main.py --audio path/to/call.mp3

Iqoqo cubungula uhla lwemibhalo:

python main.py --batch data/audio/

Ukuze uthole umuzwa ophelele wokusebenzisana:

python main.py --dashboard

Vula esipheqululini sakho.

Umdwebo 8: Okukhiphayo kwetheminali okubonisa ukuhlaziya okuyimpumelelo okunamaphuzu emizwa

# Isiphetho

Sakhe uhlelo oluphelele, olukwazi ukungaxhunyiwe ku-inthanethi oluloba izingcingo zamakhasimende, luhlaziye imizwa nemizwelo, futhi lukhiphe izihloko eziphindelelayo — konke ngamathuluzi omthombo ovulekile. Lesi isisekelo esilungele ukukhiqiza:

• Amaqembu okusekela amakhasimende ahlonza amaphuzu ezinhlungu
• Abaphathi bomkhiqizo baqoqa impendulo esikalini
• Ukusebenza kwe-ejenti eqinisekisa ikhwalithi

Ingxenye engcono kakhulu? Yonke into isebenza endaweni, ihlonipha ubumfihlo bomsebenzisi futhi isusa izindleko ze-API.

Ikhodi ephelele iyatholakala ku-GitHub: I-An-AI-ihlaziya-imizwa-yekhasimende. Vala ikhosombe, landela lesi sifundo se-AI senkulumo-kuya-umbhalo, bese uqala ukukhipha imininingwane kumakholi akho ekhasimende namuhla.

Shithu Olumide ungunjiniyela wesofthiwe nombhali wezobuchwepheshe othanda ukusebenzisa ubuchwepheshe obuphambili ekwenzeni izindaba ezithokozisayo, oneso elibukhali lemininingwane kanye nekhono lokwenza imiqondo eyinkimbinkimbi ibe lula. Ungathola futhi i-Shittu Twitter.