Sales call intelligence

1
import requests
2
import time
3
from collections import Counter
4
5
# ── Config ────────────────────────────────────────────────────
6
base_url = "https://api.assemblyai.com"
7
headers = {"authorization": "YOUR_API_KEY"}
8
9
audio_url = "https://assembly.ai/wildfires.mp3"
10
11
# ── Step 1: Transcribe with speaker labels + sentiment analysis ──
12
data = {
13
    "audio_url": audio_url,
14
    "speech_models": ["universal-3-pro"],
15
    "speaker_labels": True,
16
    "sentiment_analysis": True,
17
}
18
19
response = requests.post(base_url + "/v2/transcript", headers=headers, json=data)
20
response.raise_for_status()
21
transcript_id = response.json()["id"]
22
23
while True:
24
    result = requests.get(f"{base_url}/v2/transcript/{transcript_id}", headers=headers).json()
25
    if result["status"] == "completed":
26
        break
27
    elif result["status"] == "error":
28
        raise RuntimeError(f"Transcription failed: {result['error']}")
29
    time.sleep(3)
30
31
# ── Step 2: Identify speakers by role ──
32
understanding_response = requests.post(
33
    "https://llm-gateway.assemblyai.com/v1/understanding",
34
    headers=headers,
35
    json={
36
        "transcript_id": transcript_id,
37
        "speech_understanding": {
38
            "request": {
39
                "speaker_identification": {
40
                    "speaker_type": "role",
41
                    "known_values": ["Sales Rep", "Customer"],
42
                }
43
            }
44
        },
45
    },
46
)
47
understanding_response.raise_for_status()
48
identified = understanding_response.json()
49
50
# ── Step 3: Calculate talk/listen ratio per speaker ──
51
speaker_durations = Counter()
52
for utterance in identified["utterances"]:
53
    duration_ms = utterance["end"] - utterance["start"]
54
    speaker_durations[utterance["speaker"]] += duration_ms
55
56
total_ms = sum(speaker_durations.values())
57
talk_ratios = {
58
    speaker: round(dur / total_ms * 100, 1)
59
    for speaker, dur in speaker_durations.items()
60
}
61
62
# ── Step 4: Summarize sentiment shifts ──
63
sentiment_by_speaker = {}
64
for s in result["sentiment_analysis_results"]:
65
    speaker = s.get("speaker", "Unknown")
66
    sentiment_by_speaker.setdefault(speaker, []).append(s["sentiment"])
67
68
sentiment_summary = ""
69
for speaker, sentiments in sentiment_by_speaker.items():
70
    counts = Counter(sentiments)
71
    sentiment_summary += (
72
        f"{speaker}: "
73
        f"{counts.get('POSITIVE', 0)} positive, "
74
        f"{counts.get('NEUTRAL', 0)} neutral, "
75
        f"{counts.get('NEGATIVE', 0)} negative\n"
76
    )
77
78
# ── Step 5: Format transcript and generate coaching scorecard ──
79
speaker_transcript = "\n".join(
80
    f"{u['speaker']}: {u['text']}" for u in identified["utterances"]
81
)
82
83
llm_response = requests.post(
84
    "https://llm-gateway.assemblyai.com/v1/chat/completions",
85
    headers=headers,
86
    json={
87
        "model": "claude-sonnet-4-5-20250929",
88
        "messages": [
89
            {
90
                "role": "user",
91
                "content": (
92
                    "You are a sales coaching assistant. Analyze this sales call and produce a scorecard.\n\n"
93
                    f"Talk/listen ratios: {talk_ratios}\n\n"
94
                    f"Sentiment breakdown:\n{sentiment_summary}\n"
95
                    f"Transcript:\n{speaker_transcript}\n\n"
96
                    "Produce:\n"
97
                    "1. Call summary (2-3 sentences)\n"
98
                    "2. Talk/listen ratio analysis (ideal is 40/60 for the rep)\n"
99
                    "3. Customer sentiment shifts and what caused them\n"
100
                    "4. Top 3 coaching suggestions for the sales rep"
101
                ),
102
            }
103
        ],
104
        "max_tokens": 2000,
105
    },
106
)
107
llm_response.raise_for_status()
108
109
print("=== Sales Call Scorecard ===\n")
110
print(llm_response.json()["choices"][0]["message"]["content"])